bank marketing research paper

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• PMC10783788

Enhancing bank marketing strategies with ensemble learning: Empirical analysis

1 Institute of Traffic Engineering, Nanjing Vocational University of Industry Technology, Nanjing, Jiangsu, China

2 School of Mathematics, Sichuan University, Chengdu, Sichuan, China

Associated Data

All relevant data are within the paper and its Supporting Information files.

In order to enhance market share and competitiveness, large banks are increasingly focusing on promoting marketing strategies. However, the traditional bank marketing strategy often leads to the homogenization of customer demand, making it challenging to distinguish among various products. To address this issue, this paper presents a customer demand learning model based on financial datasets and optimizes the distribution model of bank big data channels through induction to rectify the imbalance in bank customer transaction data. By comparing the prediction models of random forest model and support vector machine (SVM), this paper analyzes the ability of the prediction model based on ensemble learning to significantly enhance the market segmentation of e-commerce banks. The empirical results reveal that the accuracy of random forest model reaches 92%, while the accuracy of SVM model reaches 87%. This indicates that the ensemble learning model has higher accuracy and forecasting ability than the single model. It enables the bank marketing system to implement targeted marketing, effectively maintain the relationship between customers and banks, and significantly improve the success probability of product marketing. Meanwhile, the marketing model based on ensemble learning has achieved a sales growth rate of 20% and improved customer satisfaction by 30%. This demonstrates that the implementation of the ensemble learning model has also significantly elevated the overall marketing level of bank e-commerce services. Therefore, this paper offers valuable academic guidance for bank marketing decision-making and holds important academic and practical significance in predicting bank customer demand and optimizing product marketing strategy.

1. Introduction

With the rapid advancement of internet and information technology, banks are increasingly diversifying their marketing strategies and expanding their efforts across both offline and online channels by utilizing various new media platforms [ 1 – 3 ]. Marketing ability is not only related to the bank’s customer acquisition and retention, but also related to its brand influence, market share and profitability. But how to effectively predict and improve their own marketing ability? At present, there are few comprehensive models or tools that can comprehensively predict the marketing ability of banks. The existing research focuses more on a single factor or local problem, lacking a global perspective. Therefore, it provides an efficient and data-driven forecasting tool for the banking industry to improve the accuracy and efficiency of marketing decisions. This is of great significance for the banking industry to succeed in the highly competitive market [ 4 , 5 ]. As China’s economy continues to grow, the number of commercial institutions in the country is on the rise. Consequently, commercial banks find themselves in increasingly fierce competition. Furthermore, the rapid expansion of online financial services is leading to a significant decline in reliance on offline banking customers, which in turn is giving rise to a growing number of customer-related challenges. One of the primary tasks facing commercial banks today is the need to maintain positive relationships with their existing offline clients while also enhancing their marketing capabilities to meet evolving customer demands. To better manage these customer relationships and navigate the challenges posed by the rapid development of the internet, data mining technology has been widely adopted across various aspects of life [ 6 ]. In marketing, banks need to accurately predict customers’ needs, behaviors and reactions to formulate relevant strategies and plans. Ensemble learning can significantly improve the prediction accuracy by combining the prediction results of several basic models [ 7 ]. This is very important for banks, because accurate forecasting can help them better locate customers and launch targeted products and services, thus improving sales efficiency and customer satisfaction. This provides strong support for the creation of a bank marketing capability model.

This paper introduces a customer demand learning model based on a financial dataset, which advances the overall effectiveness of bank e-commerce by simulating and predicting actual customer business needs. It is grounded in research on internet technology, ensemble learning models, and an analysis of traditional banking marketing models. The following outlines the paper’s primary logical structure: Section 1 provides an overview of the historical development of ensemble learning and internet technologies. Section 2 consolidates recent research on ensemble learning models and the growth of bank marketing capabilities. In Section 3, a forecasting system for a bank’s marketing capabilities is established through data analysis of bank customers. The model’s predictions and experimental findings are discussed in Section 4. Finally, Section 5 presents the research conclusions after a thorough examination of the data results. The paper is a useful source of information for encouraging the commercial banks’ marketing skills to grow intelligently.

2. Recent related work

2.1 recent work on ensemble learning model.

Mao et al. (2019) [ 8 ] investigated the maximization and diversity of ensemble learning through transformation. They developed a weighted ensemble learning technique that simultaneously maximized variety and individual accuracy, contributing to the latest advancements in ensemble learning models. In the proposed framework, several basic learners are combined and then turned into linear transformations of each of these basic learners. The optimal weight is then determined by pursuing the best projection direction of the linear transformations. The performance of the suggested method is demonstrably improved by comparison to other ways according to experimental findings. Dong et al. (2020) [ 9 ] analyzed ensemble learning using cutting-edge computer science technology and identified ensemble learning as a research hotspot. The findings demonstrated that the suggested model can successfully construct an effective knowledge discovery and mining system by integrating data fusion, data modeling, and data mining into a single framework. Alam et al. (2020) [ 10 ] proposed a novel dynamic ensemble learning algorithm based on neural network fusion after conducting a study on dynamic ensemble learning algorithms for neural networks. The experimental findings supported the notion that the dynamic neural network ensemble produced by the new dynamic ensemble learning has a suitable design, a diverse population, and a high degree of generalizability. Chen et al. (2021) [ 11 ] employed deep learning technology to investigate network security in smart cities. The research demonstrated that the proper design based on deep learning method was very important to protect smart city networks through the analysis and comparison of deep learning models such as Boltzmann machine, restricted Boltzmann machine, Deep Belief Network (DBN), Recurrent Neural Network (RNN), and convolutional neural network (CNN). Matloob et al. (2021) [ 12 ] contributed to the field by providing concise insights into the latest trends and developments in ensemble learning for software defect prediction. They also predicted and studied software defects based on ensemble learning. The performance of the model’s predictions was evaluated using the ensemble learning technique. The study demonstrated that feature selection and data sample were essential preprocessing stages that could enhance the effectiveness of an ensemble classifier. By system mapping study and cross-benchmark evaluation, Tama et al. (2021) [ 13 ] examined the ensemble learning technology of intrusion detection system and made a thorough empirical evaluation on the most recent advancements in ensemble learning technology. The findings demonstrated that by ensemble learning from intrusion detection systems, the model’s prediction accuracy was greatly increased.

Furthermore, in regards to the framework system of domain adaptive ensemble learning, Zhou et al. (2021) [ 14 ] studied the unifying framework of domain adaptive ensemble learning, and studied the non-expert ensemble learning from other sources to provide supervision signals through the analysis of multi-source unsupervised domain adaptation. Domain adaptable ensemble learning has significantly improved on these two concerns, and typically offers several benefits according to a huge number of trials. In their study of the forest fire detection system based on ensemble learning, Xu et al. (2021) [ 15 ] integrated two different learners, Yolov5 and EfficientDet, to complete the fire detection process. Tests on data sets demonstrated that the suggested strategy reduced the false alarm rate by 51.3% without adding any extra time and increased detection performance by 2.5% to 10.9%. Lin et al. (2021) [ 16 ] investigated the plug-in hybrid vehicle’s energy management method using ensemble learning speed prediction. According to the verification results, the suggested technique outperformed the benchmark strategy in terms of great fuel efficiency over a variety of driving cycles. Zhang et al. (2022) [ 17 ] researched and developed a method based on ensemble learning, studied the slope stability prediction technology based on that technology, predicted the slope stability by introducing random forest and extreme gradient lifting, and then applied that method to the slope stability prediction and research in Chongqing. The findings suggested that the ensemble learning-based proposed method offered a potential way to appropriately capture the slope state. Abbasi et al. (2022) [ 18 ] used ensemble learning to determine and study identity. Their experimental findings revealed that the ensemble learning approach suggested in the study improved accuracy by 14.2%. Yin et al. (2022) [ 19 ] conducted research on the ensemble learning model of the Bayesian optimization technique. By combining the extreme gradient lifting and random forest models in an ensemble learning approach, the XGBoost model outperformed the random forest model in terms of accuracy, precision, recall, F1 score, and kappa coefficient. The study could serve as a guide for creating maps of mineral prospects and perfecting ensemble learning models. Nguyen et al. (2022) [ 20 ] devised a computer technique to virtually screen possible anti-cancer drugs and used ensemble learning paired with evolutionary computation to identify anti-cancer natural products. They also built an ensemble computing framework through machine learning. According to the study, using the ensemble model to identify natural compounds with anticancer activity has real-world applications.

The rapid development of the Internet of Things (IoT) and information technology (IT) has highlighted the inadequacy of single data classifiers in handling the demands of big data processing. Ensemble learning technologies address this challenge by employing multiple data classifiers, enabling stacking, additional processing, and prediction generation from model data. This approach not only enhances prediction accuracy but also increases the effectiveness and scalability of data processing.

2.2 Progress and related work on bank marketing ability

The potential sources of climate funding in underdeveloped nations were compared by Banga (2019) [ 21 ] with regard to the assessment and analysis of banks’ marketing capabilities. The findings of the study on banks’ marketing capabilities and green bonds demonstrated that, thanks to investors’ growing climate consciousness, green bonds were on the increase in both developed and emerging markets. However, in developing nations, the green bond market was still in its infancy, and its full potential had not been fully realized. Boating (2019) [ 22 ] investigated the link between online relationship marketing and customer loyalty, collected experimental data by looking into 429 Ghanaian retail bank clients, and used structural equation modeling technology to assess the findings. The study demonstrated that in addition to the online technologies employed, banks’ online relationship activities must also transmit pertinent and beneficial signals to support the promotion of their marketing skills. In order to gather information, Nazaritehrani et al. (2020) [ 23 ] researched the growth and market share of e-banking channels in developing nations and created a questionnaire. The expert judgments and the Kronbach technique were used to evaluate the scale’s validity and reliability. The study noted that although there was a statistically significant association between a bank’s market share and the growth of its marketing capabilities, the outcome was not significant. Asnawi et al. (2020) [ 24 ] investigated the significance of customer satisfaction and loyalty in Indonesia and its impact on Islamic banks, testing the theory by compiling information from 280 Indonesian bank clients. Managers can specifically tailor their services for Muslim customers by evaluating the Islamic banks’ customer service standards. The gap between what customers expect from the services offered and what they actually receive can be reduced and customer satisfaction can also be increased while establishing the bank’s marketing strategy. Mulia et al. (2021) [ 25 ] investigated the effect of customer intimacy in enhancing customer loyalty of Islamic banks, and the information was gathered through self-management investigation techniques. Meanwhile, data analysis employed multivariate linear regression and multivariate analysis of variance. The findings indicated that loyalty was impacted both directly and indirectly by customer closeness.

Several factors influence a bank’s effectiveness in marketing itself within the financial and commercial markets. Arjun et al. (2021) [ 26 ] conducted research into the evolving landscape of bank intelligence in emerging countries, focusing on the distinct intelligent decision support model of the banking industry. Their study examined critical elements that impact a bank’s marketing capability and demonstrated how the advancement of artificial intelligence (AI) and Internet of Things (IoT) technologies could enhance these capabilities. Rahmayati (2021) [ 27 ] empirically analyzed the competitive strategy of Islamic banking, gathering data from written sources. The study’s outcomes revealed that banks can increase customer satisfaction through the implementation of proactive marketing methods. Boustani (2022) [ 28 ] investigated the impact of AI on customers and staff in banks across developing Asian nations. Using a model based on quantitative research and a fictitious regression model for analysis, the study found that AI elevated the quality of bank transactions, serving as a practical benchmark for predicting and assessing a bank’s marketing potential. In order to get experimental data, Mogaji et al. (2022) [ 29 ] conducted semi-structured interviews with 47 bank managers from both developed and developing nations. They researched the AI technology connected to marketing financial services. The conceptual framework of AI connected to financial service marketing presented in this paper captured and emphasized the interaction between clients, banks, and outside stakeholders. The research was a useful source of information for enhancing bank marketing capabilities. With regard to Indian banking, Sharma et al. (2022) [ 30 ] conducted qualitative and quantitative research, proposed a conceptual model for the green banking effort, and examined the impact of the green banking initiative on bank marketing prowess. The study demonstrated that the green bank strategy contributed favorably to improving the green image and regaining client trust. Galletta et al. (2022) [ 31 ] used samples from 205 Nordic listed businesses during the period of 2002 to 2020 to investigate gender diversity and sustainable performance in the banking sector, as well as the relationship between CSR and environmental management. The research showed that the correlation between gender diversity of board of directors and sustainable performance is more obvious in the carbon-intensive industry sub-sample.

In conclusion, the era of big data has ushered in changes in the marketing model, gradually replacing the traditional telemarketing model with the internet marketing model due to the substantial increase in data volume. To better understand the evolution of commercial banks’ marketing capabilities and identify the algorithm models best suited for accurate marketing of bank wealth management products, further analysis of the user structure of these products is necessary. This analysis is crucial for advancing business operational methods.

3. Prediction and evaluation model of bank marketing big data distribution based on ensemble learning

3.1 ensemble learning model and bank marketing.

As IoT technology has advanced, the precision of machine learning tasks has become more crucial than their processing speed [ 32 ]. By developing a robust learning model structure within the ensemble learning framework, machine learning tasks can be effectively completed. In assessing participants’ marketing skills based on data, the expected value of each model evaluator generates a random combination, typically resulting in more precise predictions than those of individual participants. A hierarchical control decision-making structure model for learning monitoring is created to handle classification, regression, and preprocessing tasks in advanced ensemble learning concepts. When the loss function is straightforward, the data optimization phase can be relatively straightforward, involving the error quadrant function or exponential function for data processing. The ensemble learning model’s data interaction mechanism among the user layer, server layer, and data interaction layer is examined. Fig 1 illustrates the system architecture of the ensemble learning model.

In Fig 1 , in the context of bank marketing ability prediction, customers can be the marketing team, analysts, business decision makers or other relevant personnel of the bank. Customers can connect with the data interaction layer through applications, user interfaces or other means, and request forecasts or other relevant information from the server. The data interaction layer is a key component to connect the client and the server. It plays a bridge role in data transmission, communication and interaction. A server is a computer system or cloud service that stores and hosts an ensemble learning model. It is responsible for receiving requests from customers, and then calling or running corresponding models to generate prediction results.

3.2 Bank customer demand and financial big data collection and analysis

In the banking industry, accurately predicting whether customers will make time deposits is of utmost importance when analyzing real customer data through the banking marketing system [ 33 ]. Big data collection is an Extract-Transform-Load (ETL) operation on data sources. When collecting financial data, due to the complexity of data sources, before data analysis, it is necessary to extract the data needed for precise marketing from the complex original data format through extraction technology.

The financial data extraction in this paper is a process of extracting the required data from the big data of existing marketing systems in banks (China Merchants Bank, China Construction Bank, China Agricultural Bank and China Industrial and Commercial Bank) through ETL tools. Data extraction is summarized according to the agreed XML format and Service format. First, it is filtered and analyzed, and then the data is transformed into information and knowledge. The methods of data extraction include total extraction and incremental extraction. Total extraction is equivalent to data migration or copying, which extracts the data from the table or view in the data source from the database and converts it into the format recognized by ETL tools. Incremental extraction refers to extracting new or modified data from the database. Incremental extraction is more widely used in ETL tools than full extraction of data, and how to capture changing data is the key to incremental extraction of data. Firstly, the total data of 2020 and 2021 are crawled from different subsystems and stored in the data warehouse. Meanwhile, ETL tools extract customer data, sales data and regional data needed for precise marketing from the big data of the existing marketing system of banks. Then, in the end, this paper extracts a total of 25317 bank sales data through the above extraction methods. When collecting and using financial data, it strictly abides by ethical and privacy standards, respect the confidentiality and sensitivity of financial data, and ensures the legal acquisition and use of data.

The process of model interpretable analysis begins with data analysis. Firstly, the research requires a comprehensive understanding of the data, encompassing an assessment of the overall data distribution and an analysis of feature relationships. The second step involves the use of data visualization techniques to meticulously characterize and analyze the data before constructing a data-driven self-service framework. By considering both the correlations within the dataset and those between data elements and target variables, uncovering hidden patterns within financial datasets can significantly enhance the model’s predictive capacity. This optimization of information data for the features required by machine learning models holds substantial reference value. Fig 2 illustrates the structure of the customer data analysis and management system, providing an overview of the server-side structure and management framework for bank marketing businesses.

In Fig 2 , in this structure diagram, the client is connected to the bank and the management server through the network, and can perform various data analysis and management tasks. The bank is responsible for data analysis and model training, while the management server is used for system configuration and monitoring. Server nodes provide computing and storage resources to support the operation of the system. Finally, the management server provides centralized management and control of the system to ensure the stability and maintainability of the system.

3.3 Analysis on the construction of bank marketing ability forecasting model based on ensemble learning

Ensemble Learning, as one of the machine learning algorithms, completes the learning task by combining multiple individual learners. These learners have the same type and different types. The same type is called homogeneous and the different types are called heterogeneous [ 34 ]. Usually, homogeneous algorithms usually include Boosting algorithm and Bagging algorithm, while heterogeneous algorithms combine different types of classifiers by adopting some fusion strategy, such as Stacking algorithm. Although Stacking algorithm combines the advantages of multiple models, the calculation cost of Stacking is high, and cross-validation and the construction of multiple base models make it consume a lot of calculation resources and time. Therefore, this paper mainly analyzes lifting method and bagging method.

The advantage of Boosting algorithm is that it can combine different learners with weights, thus improving the performance of the model. At present, the widely used Boosting algorithms include GBDT algorithm, XGBoost algorithm and LightGBM algorithm. Among them, Extreme Gradient Boosting (XG Boost) is a machine learning system. It adopts distributed gradient lifting algorithm, which can effectively improve the accuracy and efficiency of classification, regression and feature selection. The objective function is shown in Eq ( 1 ):

L refers to the loss function and Ω refers to the positive term. For a tree containing ( t -1) integration trees, the ( t )th feature decision tree can be expressed as Eq ( 2 ):

y ^ l ( t − 1 ) refers to the predicted value given by the model in step (t-1), it is a known constant. f i ( x i ) refers to the predicted value of the feature tree to be added this time. In this case, the objective function is shown in Eq ( 3 ).

According to Taylor formula, the function f ( x +Δ x ) is expanded in Taylor’s second order at point x , and the Eq ( 4 ) can be obtained:

Taylor’ formula is brought into the objective function of XGBoost, where x corresponds to the predicted value y ^ l ( t − 1 ) of the first ( t -1) feature tree and Δ x corresponds to the t tree f i ( x i ) being trained, and the objective function is rewritten as Eq ( 5 ):

The objective function obtained by removing the constant term is shown in Eq ( 6 ).

Compared with Boosting algorithm, Bagging algorithm aims at reducing the variance of the model, that is, reducing the sensitivity of the model to data, thus improving the generalization ability of the model. Random forest introduces more randomness on the basis of Bagging algorithm. By randomly selecting some features every time a node is split, the correlation of the model is reduced and the diversity of the model is improved. Therefore, random forest is more suitable for data sets that are unstable or susceptible to noise. Therefore, random forest is more suitable for data sets that are unstable or susceptible to noise.

Suppose a random forest model contains T decision trees, and the test sample x needs to be classified into category y . Assuming that each decision tree is independent of each other, the whole random forest can be regarded as the average of multiple decision tree classifiers, as shown in Eq ( 7 ):

f i ( x ) refers to the classification result of sample x by the ith decision tree. Let the variance of each decision tree be σ i 2 , then the variance of the whole random forest is shown in Eq ( 8 ):

C o v ( f i ( x ) , f j ( x ) ) can be expressed as Eq ( 9 ):

Because the classification prediction results of each decision tree are independent of each other, there exists Eq ( 10 ).

When i and j take different values, E [ ( f i ( x ) , f j ( x ) ) ] is shown in Eq ( 11 ):

Therefore, the variance of random forest can be rewritten as Eq ( 12 ):

Among them, the first term is the variance of a single decision tree, and the second term is the covariance between decision trees. Because of the differences between decision trees, the covariance is often small and can usually be ignored. The above equation can be approximated as Eq ( 13 ):

The variance of random forest is approximately equal to the variance of a single decision tree divided by the number of decision trees. It shows that random forest can greatly reduce the variance of model prediction results and obtain better classification accuracy.

In order to improve the algorithm, this paper combines XGBoost algorithm in ensemble learning with random forest algorithm and applies it to the prediction of bank marketing ability. Fig 3 describes the forecasting model of bank marketing ability based on ensemble learning.

In Fig 3 , the model firstly uses the random forest algorithm to make feature selection and preliminary prediction for the bank marketing system, and then takes the output of the random forest as the input feature of XGBoost, and uses XGBoost to train a new model to better capture the complex relationships and patterns in the data. Finally, the forecast results of bank marketing ability are integrated and output.

In this model, accurate analysis of bank marketing opportunities typically uses superposition, under sampling, and thorough sampling methods. By increasing the number of samples, overlapping sampling corrects for positive sampling, but this necessarily results in a huge amount of noise data [ 35 ]. By reducing the proportion of most samples, the primary tenet of an under-sampling technique is to enhance the classification effect of the model in a few categories. Regrettably, this approach may slightly reduce the model’s accuracy prediction rate while also reducing the sample size. These two sampling techniques are combined to create a comprehensive sampling strategy. Overlapping sampling involves adding a small number of samples to the training set, while under sampling removes more incorrect sample data. Extensive grid search for multiple parameters could potentially impact the model’s performance. The random forest algorithm must be utilized to look up machine learning parameters in order to resolve this issue. It is required to input the weight of participants in the model to increase the marketing forecasting capability of banks in accordance with the accuracy of the prediction results of the basic model because the parameter accuracy of each ensemble learning model differs. User management, project management, and sales management make up the bulk of the network system for managing bank marketing capability. Fig 4 depicts the precise structural framework.

In Fig 4 , the core part of the network architecture includes user management, project management and sales management, which work together to help banks better organize and coordinate marketing activities, improve marketing capabilities, meet customer needs and achieve sustainable growth.

3.4 Experimental evaluation

When analyzing the performance of the bank marketing ability prediction model based on ensemble learning, the input data of the model must be cleaned and preprocessed first to obtain appropriate prediction results. The data processing process of the server and the model is analyzed, and the data processing process of the established initial learner is shown in Fig 5 .

In Fig 5 , the model’s input data must first be cleaned and preprocessed for the model to obtain appropriate prediction results. The first steps involve filling in the gaps in the data, cleaning out the noise, and addressing the imbalance between the positive and negative sample data. Second, in order to create a new combination of structural features, the original features of the data must be processed and synthesized. Finally, using a suitable algorithm model, feature extraction and model construction are carried out.

When evaluating the performance of the model, a comparative experiment is adopted. In this process, the data set is divided into training set and test set according to the ratio of 7:3, and the basic learner is cross-verified by 50%. In the model, the super parameter optimization of XGBoost algorithm is as follows: max_ depth is 4, gamma is 0.7, min_ child- weight is 5, Reg_ alpha is 0.01, reg_lambda is 100, and num_leaves is 15.

In contrast experiments, random forest, as an ensemble learning method, can reduce the variance of the model and improve the generalization performance by constructing multiple decision trees and combining their prediction results. Decision tree, Support Vector Machine (SVM) and K-nearest neighbor (KNN) are one of the machine learning algorithms respectively, and ensemble learning also belongs to the machine learning algorithm. Therefore, the model algorithm proposed in this paper is compared with the model algorithm proposed by decision tree, random forest, SVM, KNN and Yin et al. (2022), and evaluated from the indicators of accuracy, precision, recall rate, F1 value, data transmission delay and overall system delay. Among them, the accuracy of all experiments is the accuracy of 10 cross-validation under the optimal parameters, and the dispersion degree of accuracy is reflected by calculating the standard deviation of data.

The model’s performance evaluation findings demonstrate the superiority of the ensemble learning-based marketing ability prediction model. The bank marketing system may conduct targeted marketing for consumers and strengthen the relationship between customers and banks by contrasting the prediction models of random forest and SVM as well as contrasting the model suggested in this paper with these models.

4. Results and discussion

4.1 prediction performance of different types of bank marketing ability prediction models.

The performance of the ensemble learning model proposed in this paper is compared with the performance of Decision tree, Random Forest, SVM, KNN and the model algorithm proposed by Yin et al. (2022). The change trend of model prediction accuracy performance is shown in Fig 6 . This comparison is done to assess the overall prediction performance of various types of bank marketing ability prediction models. In addition, Figs ​ Figs7 7 – 9 demonstrate the analysis model’s prediction precision rate, recall rate, and F1 value trend, respectively.

Fig 6 shows that with the increase of model iterations, the prediction accuracy of various bank marketing ability prediction models first rises rapidly, and then tends to be stable. When the model is iterated for 100 times, the accuracy of each model tends to be stable, in which the prediction accuracy of decision tree model is 37%, that of SVM model is 49%, and that of suggested model is 85%. After 250 iterations, the prediction accuracy of SVM model is 65%, while the prediction accuracy of this model algorithm is 91%. The results show that the prediction accuracy of this model is higher than other models.

Fig 7 shows that the prediction precision of different types of bank marketing ability prediction models increases with the increase of model iterations. When the number of model iterations is 100, the prediction precision of SVM model is 38%, while that of KNN model is 45%, and that of ensemble learning model is 61%. After 250 iterations, the prediction precision of SVM model, KNN model and ensemble learning model is 41%, 49% and 69% respectively. By comparing the prediction precision curves of different models, the performance of this model is better than other models.

In Fig 8 , when the number of model iterations is increased to 100, the change data of the predicted recall rate of several different types of bank marketing ability prediction models are constantly rising. When the model is iterated for 100 times, the prediction recall rate of the bank marketing ability evaluation model based on SVM is 44%, while the prediction recall rate of the bank marketing ability evaluation model based on KNN is 54%, and the prediction recall rate of the K-means model is 64%. The prediction recall rate of the model based on ensemble learning proposed in this paper is 73%. When the number of iterations of the model exceeds 100, the recall value tends to be basically stable. When the number of model iterations is 250, the prediction recall rate of the model based on ensemble learning is 74%. Therefore, the comprehensive performance of this model is better than other models.

In Fig 9 , when the number of model iterations increases to 100 times, the predicted F1 value data of the whole system shows an overall growth trend. When the model is iterated for 100 times, the predicted F1 value of the banking marketing system based on SVM is 41%, while the predicted F1 value of the model based on KNN is 49%, and the predicted F1 value of the ensemble learning model proposed by the research can reach 75%. However, when the number of iterations of each model is 100 to 250 times, the F1 value of each model algorithm tends to be basically stable. When the iteration number of the model is 250, the predicted F1 value of the banking marketing system based on SVM is 50%, while the predicted F1 value based on the ensemble learning model proposed in this paper is 74%.

Through the analysis of the above results from accuracy, precision, recall and F1value, it is found that the prediction accuracy of the model algorithm in this paper is up to 91%. This may be due to the in-depth understanding of the characteristics of banking marketing in this model, and the integration of XGBoost algorithm and random forest algorithm in ensemble learning can capture the nonlinear relationship or highly complex patterns in the data, so the model algorithm in this paper can improve the generalization ability of the model and obtain the optimal prediction accuracy. However, random forest and SVM are different in data distribution, parameter adjustment and category imbalance, so they show different accuracy.

4.2 System data transmission performance of different types of models

The data transmission performance of different types of models is analyzed. With the change of model iterations, the data transmission delay curves of different types of bank marketing ability prediction models are shown in Fig 10 . In addition, the numerical change curves of the overall network delay of different types of bank marketing ability prediction models are shown in Fig 11 .

In Fig 10 , in different types of bank marketing ability prediction models, the data transmission delay of different models fluctuates with the increase of iteration times. When the number of model iterations is 50, the data transmission delay of the prediction model based on decision tree is 60s, that of the prediction model based on SVM is 46s, and that of the proposed model is only 14s. After 250 iterations of the model, the data transmission delay of the prediction model based on SVM is 52s, while the data transmission delay of the prediction system based on ensemble learning model proposed in this paper is only 30s.

In Fig 11 , with the increasing number of network iterations, the overall network data transmission delay of the model is increasing. When the iteration number of the model is 50, the system transmission delay of the banking marketing forecasting system based on SVM is 42s, while the network transmission delay of the banking marketing system based on KNN algorithm is 32s. At this time, the network data transmission delay of K-means banking marketing system is 23s, while the network data delay of the proposed banking marketing system is only 13s. After 250 model iterations, the system transmission delay of the banking marketing forecasting system based on SVM is 54s, while the whole network transmission delay of the forecasting system based on the ensemble learning model proposed in this paper is only 27s. Comparing the comprehensive performance of different models, the overall data transmission performance of the bank marketing system based on ensemble learning model is better than other models.

Further, the model algorithm in this paper is compared with other model algorithms in terms of sales growth rate and customer satisfaction, and the results are shown in Table 1 .

In Table 1 , the model algorithm in this paper shows the highest results in sales growth rate and customer satisfaction growth rate, which are 25.67% and 20.52% respectively. In contrast, the results of other algorithms are lower than the model algorithm in this paper, among which Yin et al. (2022) takes the second place in terms of sales growth rate and customer satisfaction growth rate, which are 21.36% and 18.74% respectively. Therefore, the relatively good effect of this model in the prediction of bank marketing ability can help banks better understand and improve their marketing ability.

4.3 Discussion

The results of this paper show that the proposed algorithm has achieved remarkable success in forecasting and improving the marketing ability of banks. Meanwhile, through this paper, its impact on bank marketing strategy is as follows:

First, the effectiveness of marketing ability prediction is improved. The algorithm in this paper shows excellent ability in forecasting the growth rate of sales and customer satisfaction. High sales growth rate and customer satisfaction growth rate show that the algorithm can effectively predict which marketing strategies and activities may be successful and have a positive impact on customers, which is consistent with Sugiato et al. (2023) [ 36 ]. This helps banks allocate resources and energy wisely to achieve better business performance.

Second, the accuracy of marketing decisions is improved. By using this algorithm, banks can predict the effect of marketing activities more accurately to make better decisions. Accurate marketing forecast helps to avoid the waste of resources, optimize advertising and promotion activities, and better meet customer needs. This is expected to improve the market competitiveness of banks.

Third, the efficiency of marketing decisions is improved. Through accurate forecasting, banks can make decisions more quickly, reduce the cost of trial and error, and adjust market strategies more flexibly. This helps banks adapt to market changes more quickly and seize business opportunities.

5. Conclusion

New marketing strategies, particularly those leveraging internet technologies, have significantly transformed the banking marketing landscape with the emergence of e-commerce technologies. This paper aims to delve into the data interaction process of the ensemble learning model through the utilization of machine learning and ensemble learning techniques. Based on the analysis of the traditional marketing model of banks, this paper builds an automatic learning model to analyze and predict customer data. The bank marketing system may carry out targeted marketing for customers, further sustain the relationship between customers and banks, and increase the likelihood of successful product marketing by contrasting the prediction models of random forest and SVM. The findings highlight the potential to establish a differentiated banking product marketing system using a forecasting model based on ensemble learning, further enhancing the market segmentation strategy of e-commerce banks. However, this paper also has some shortcomings. If the time span of data collection may be short, the data in a longer time range can be considered in future research to enhance the prediction performance of the model. In the future, in order to further improve the forecasting effect and marketing ability of the model, it is necessary to modify the model parameters through grid search in the future research. Moreover, the accuracy and practicability of the model can be further improved by further exploring the field of bank marketing ability prediction.

Supporting information

Funding statement.

The authors received no specific funding for this work.

Data Availability

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here .

Loading metrics

Open Access

Peer-reviewed

Research Article

A machine learning approach to the digitalization of bank customers: Evidence from random and causal forests

Roles Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing

* E-mail: [email protected]

Affiliations University of Granada, Granada, Spain, Bangor University, Hen Goleg, Bangor, United Kingdom, Funcas, Madrid, Spain, CUNEF, Madrid, Spain

Roles Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing

Affiliations Funcas, Madrid, Spain, CUNEF, Madrid, Spain

Affiliations University of Granada, Granada, Spain, Funcas, Madrid, Spain

• Santiago Carbo-Valverde, 
• Pedro Cuadros-Solas, 
• Francisco Rodríguez-Fernández

• Published: October 28, 2020
• https://doi.org/10.1371/journal.pone.0240362
• Reader Comments

Understanding the digital jump of bank customers is key to design strategies to bring on board and keep online users, as well as to explain the increasing competition from new providers of financial services (such as BigTech and FinTech). This paper employs a machine learning approach to examine the digitalization process of bank customers using a comprehensive consumer finance survey. By employing a set of algorithms (random forests, conditional inference trees and causal forests) this paper identities the features predicting bank customers’ digitalization process, illustrates the sequence of consumers’ decision-making actions and explores the existence of causal relationships in the digitalization process. Random forests are found to provide the highest performance–they accurately predict 88.41% of bank customers’ online banking adoption and usage decisions. We find that the adoption of digital banking services begins with information-based services (e.g., checking account balance), conditional on the awareness of the range of online services by customers, and then is followed by transactional services (e.g., online/mobile money transfer). The diversification of the use of online channels is explained by the consciousness about the range of services available and the safety perception. A certain degree of complementarity between bank and non-bank digital channels is also found. The treatment effect estimations of the causal forest algorithms confirm causality of the identified explanatory factors. These results suggest that banks should address the digital transformation of their customers by segmenting them according to their revealed preferences and offering them personalized digital services. Additionally, policymakers should promote financial digitalization, designing policies oriented towards making consumers aware of the range of online services available.

Citation: Carbo-Valverde S, Cuadros-Solas P, Rodríguez-Fernández F (2020) A machine learning approach to the digitalization of bank customers: Evidence from random and causal forests. PLoS ONE 15(10): e0240362. https://doi.org/10.1371/journal.pone.0240362

Editor: Baogui Xin, Shandong University of Science and Technology, CHINA

Received: April 8, 2020; Accepted: September 24, 2020; Published: October 28, 2020

Copyright: © 2020 Carbo-Valverde et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: The data can be accessed by request to Funcas Foundation. Data are made available upon request -not shared publicly – because data are owned by Funcas (a third-party organization). Funcas is a non-profit organization focused on fostering academic research that aims to ensure that data are solely used for academic purposes. Then, in order to ensure that data are not going to be used for non-academic purposes (e.g. marketing or commercial) Funcas’s Ethics committee requires to verify that researchers are going to use them for only for academic purposes. Authors are not responsible for ensuring data access. Data access is granted by request to Funcas Foundation (Caballero de Gracia, 28 28013 Madrid (Spain) or directly via email to [email protected] ).

Funding: Financial support from the FUNCAS Foundation, PGC2018 – 099415 – B – 100 MICINN/FEDER/UE, and Junta de Andalucía P18-RT-3571 Project and P12.SEJ.2463 (Excellence Groups) is gratefully acknowledged. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Competing interests: The authors have declared that no competing interests exist.

1. Introduction

At the end of 2019, 53.6% of the global population, or 4.1 billion people, used online digital devices, according to Information and Communications Technology (ICT). Digitalization is changing the shape of many industries and the way companies and clients interact. This digital revolution has been particularly relevant in the banking industry where the use of digital banking (online and mobile) has become one of the most strategic channels used by bank customers. The Organization for Economic Co-operation and Development (OECD) has identified some of the core properties and crosscutting effects of the digital transformation [ 1 ] as the most important business challenge currently underway. Furthermore, the OECD recognizes banking as one of the sectors where such transformation is more relevant in economic, organizational, and social terms.

On the supply side, financial institutions have gradually reacted to these changes. Banks are particularly sensitive to the transformation of information systems, the treatment of personal data, and the emergence of new (fully digital) competitors and delivery channels. Despite incorporating online distribution channels two decades ago, and in spite of the renewed digitalization wave, banks continue to develop more information and systems-oriented business models. Digitalization is not only focused on cost savings, but also includes process improvements to enhance customer experiences [ 2 – 5 ]. This effort is driven by both rival precedence [ 6 , 7 ] and changes in demand [ 8 ].

A large number of studies on banking organization and technology have addressed the adoption of the most basic electronic banking services developed over the last few decades including debit and credit cards and more recently online banking (although partially covered). Prior literature has revealed a variety of mechanisms—motivations, attitudes, behavioral intention, social systems, and associations—involved in technology adoption. These studies have found that perceived security, usefulness, quality, and convenience drive consumer adoption of online services [ 9 – 14 ]. However, the relevance of each of these factors depends on the stage of the adoption. This is an important lesson for new digital services given the heterogeneous penetration they have both geographically and demographically [ 15 ]. This is particularly relevant considering that socio-demographic characteristics—age, gender, income, and location—[ 11 , 16 , 17 ], cultural characteristics [ 18 ], and customer experience (with other products with varying levels of technological sophistication) are strongly related to the demand for online banking services [ 19 ].

However, while the initial adoption of digital services could be examined using standard parametric statistical methods, examining customers’ digital journey is more complex. Digitalization is a challenging endeavor where several factors drive digitalization decisions [ 5 , 20 ]. Machine learning methods have emerged as powerful tools for data mining [ 21 – 24 ]. Instead of being limited to making strong assumptions about the structure of the data, machine learning allows researchers to identify and display complex patterns in a data-driven form [ 25 ]. In this sense, a machine learning approach is gaining ground in examining consumer behavior such as consumer preferences for technology products Chen, Honda, & Yang [ 26 ], travel choices [ 27 ] or to model consumer response [ 21 ].

This paper aims to benefit from the advantages of following a machine learning approach in order to examine the bank customers’ digitalization process. The use of machine and causal machine algorithms in our research context allows us to reveal the process that individuals follow to make their financial digitalization choices. Unlike prior studies, we are not focused on a single dimension of the digitalization process but on several dimensions (adoption, diversity of use and bank and non-bank’s payment choices).

Methodologically, instead of ex-ante selecting a machine learning technique, we consider a number of machine learning techniques that have proved their value in this field: random forest, extreme gradient boosting, k-nearest neighbor, support vector machine, Bayesian networks and extreme learning machine (see among others [ 28 – 32 ]). After selecting the machine learning with the best performance (in terms of predicted accuracy) we use this algorithm to identify the main features predicting bank customers’ digitalization process. Then, we build a set of classification trees to illustrate the sequence of consumers’ decision-making, and, finally, we make use of causal forests (a causal machine learning technique) to estimate the existence of causal relationships in the digitalization process.

The empirical analysis relies on extensive data collected from a survey—following the structure of the Survey of Consumer Payment Choice (SCPC) [ 33 ]—about digital banking and payment services responded by 3,005 consumers between the ages of 18 and 75. This dataset allows us to explore financial digitalization in a developed country with deep internet penetration (84.6% of adults are internet users), a highly banked population (97.2% of adults have a bank account), and a growing use of electronic banking among consumers (62% of the sample individuals are e-banking users to some extent, although the degree and scope of the adoption varies substantially across individuals), according to OECD, World Bank and GlobalWeb data.

By way of preview, we find that the random forest algorithm achieves the best performance in terms of accuracy to predict bank customers’ digitalization. This algorithm -coupled with the classification trees- reveals that bank customers need to become familiar with the information content of digital services before they begin to make financial transactions. Going digital begins with information-based services and is then followed by transactional services. Customers check their bank balances, make inquiries, and explore the possibilities of the digital channels before making payments, transferring money, or engaging in other transactional services. As for the scope of digitalization, the perceived safety of digital bank services by consumers becomes a critical filter for consumers’ diversified use of digital bank services. However, there appear to be notable exceptions. In the case of mobile banking, for example, even if perceived safety influences consumers’ adoption decisions, the speed and ease of use of the device appear to be more decisive. The efficiency of this service contrasts with the adoption process of more traditional and more established bank services such as credit and debit cards, which are used on a regular basis only when they are perceived as safe and relatively costless. Moreover, consumers adopt other non-bank digital financial services (e.g., Amazon or PayPal) only after they have already become frequent and diversified digital bank customers. These results are also confirmed when using the extreme gradient boosting algorithm and plotting a Bayesian network for each of the dimensions considered. Causal forests reveal that checking online balances has the largest effect on adopting online banking, while making money transfers with a smartphone seems to be relatively more important to become a diversified mobile banking customer. Regarding the use of bank payment methods, we find that the perception of safety has the largest impact on using credit cards while the perception of cost and convenience have the largest impact on paying with debit cards.

These results seem to have relevant business implications for the banking industry when designing strategies to bring on board and keep digital users (e.g., offer digital services focused on satisfying customers’ needs), to face the increasing competition in the payment sector by BigTech and FinTech (e.g. link payments experiences with social media) or to succeed with their digitalization programs (e.g. segmenting customers). Moreover, these results are also valuable for policymakers to design efficient measures to promote financial digitalization.

The remainder of the paper is organized as follows: Section 2 reviews the related literature; Section 3 describes the dataset and the methodology employed; Section 4 discusses the main empirical results; Section 5 addresses the causal impact using causal forests; Section 6 shows the consistency of the findings over alternative supply-side explanations and presents the implications, limitations, and scope for future research; and Section 7 concludes.

2. Related works

The main relevant studies related to financial technology adoption in the digital age refer to firm management and information systems. A number of theories aim to explain the evolution of these new technologies and the interaction between the consumer and the firm. Among them, the technology acceptance model (TAM) [ 34 ] and its latter versions (TAM2 and TAM3) have become popular for explaining how people accept and adopt new technology in the context of banking. The TAM model, which is based on the theory of reasonable action (TRA) [ 35 ] and the theory of planned behavior (TPB) (Ajzen [ 36 , 37 ]), suggests that technological adoption depends on customers’ perception of the utility and ease of use of the technology. Other theories such as the diffusion of innovations (DIT) [ 38 ], the task-technology fit (TTF) [ 39 ], the unified theory of acceptance and use of technology (UTAUT) [ 40 ], and the technology resistance theory (TRT) [ 41 ] have complemented the drivers of online adoption. These theories have thereby given prominence to a number of technological components of the service and not just to consumers’ perceptions. However, as it has recently been argued, those factors explored by the existing literature on information systems may not be sufficient to explain banking digitalization [ 20 , 42 ].

From an empirical standpoint, prior studies on customers’ perceptions have identified the main factors that explain the adoption and utilization of online banking. These include security [ 9 , 10 , 13 , 43 ], ease of use [ 12 – 14 , 44 , 45 ], convenience [ 12 , 13 ], and cost [ 11 , 46 ]. Overall, consumers use e-banking services when they perceive them as safe, useful, convenient, and relatively costless. As for the relative importance of these factors, Hoehle et al. [ 10 ] have surveyed the literature and concluded that security is a major determinant of consumers’ use of e-banking services. Additionally, many of these studies highlight that a range of socio-demographic characteristics [ 11 , 16 , 17 ] and cultural characteristics [ 18 ] also influence the adoption of online banking services. Specifically, young people who have a higher income and live in areas of high internet penetration [ 11 , 47 , 48 ] are prone to using online services. However, as Montazemi and Qahri-Saremi [ 15 ] have highlighted, the importance of these socio-demographic factors depends on the stage of the adoption of online banking services within each market segment or jurisdiction. Moreover, Szopiński [ 19 ] has found that having other banking products such as mortgages and credit cards also has a significant influence on consumers’ use of online banking services.

Closely related to online banking, studies on mobile banking adoption have also recently emerged. The empirical and theoretical approaches in these studies are similar to those to online banking [ 49 – 53 ]. The results of these studies suggest that age is the most decisive factor in mobile banking adoption. However, other determinants such as trust in the device, security, and cost have also been reported to strongly influence the adoption of mobile payments [ 54 ].

Our paper aims to offer a twofold contribution to the existing literature on bank customers’ digitalization. First, by employing a machine learning approach, it reveals the patterns driving the digitalization process. Second, unlike prior studies we do not focus on a single dimension of digitalization. We explore the digital journey of bank customers by examining a number of dimensions (adoption, diversity of use and bank and non-bank’s payment choices) to provide a more complete picture of the digitalization process.

3.1 The survey

The primary data for this study were collected from a consumer survey that was specifically conducted for this research by IMOP Insights during November and December 2016. The survey participants—a population of Spanish consumers between the ages of 18 and 75—were asked about their digital preferences and in particular about those related to banking and payment services. The survey followed the structure of the Survey of Consumer Payment Choice (SCPC) originally conducted by the Federal Reserve Bank of Boston and it is currently conducted by the Federal Reserve Bank of Atlanta. However, our survey incorporated comprehensive information about consumers’ digital preferences and not just about payment services. Controlled quotas for a representative sample of the population were established based on age, sex, and location. The survey was conducted via telephone interviews and resulted in a sample size of 3,005 consumers. The human participation in this study is simply the voluntary participation of subjects in a telephone survey conducted with all the legal and sociological guarantees. The consent of all the survey participants was informed before conducting the questionnaire and this consent was documented as part of the recorded telephone survey. Data were analyzed anonymously by the authors. S1 Appendix in S1 File offers detailed information about the survey and the data collection process and all the variables extracted from the survey questionnaire.

Spain seems to be a good laboratory for this study because it has overcome the initial implementation phase of electronic banking and ranks third in the world for annual growth in mobile banking adoption, according to OECD statistics. The penetration of online banking and the general level of financial digitalization in Spanish society are similar to those in other developed economies. Consequently, the main findings—with the necessary caveats—could likely to be extrapolated to other jurisdictions or would at least be useful for informing other research in different countries. Table 1 illustrates the sample demographics. The representativeness of the survey data is assured by comparing the sample breakdown with the Spanish National Statistics Institute (INE).

• PPT PowerPoint slide
• PNG larger image
• TIFF original image

https://doi.org/10.1371/journal.pone.0240362.t001

3.2 Descriptive statistics

Fig 1 illustrates the degree to which consumers use various financial services. On average, each banking client has two bank accounts and operates with more than one entity. It is worth noting that while 79.6% of respondents have an online bank account, only 13% are exclusively online users. Regarding the type of financial activities conducted online, internet users reported accessing online banking services to check account balances (68.72% of respondents), to receive online communications from their bank (52.18%), and to make payments or transfer money (51.13%). In the case of mobile banking, the activities lean even more toward checking and communication rather than transactional services.

Fig 1 illustrates the degree to which consumers use various financial services: receiving emails, paying bills, making payments and checking the account balances.

https://doi.org/10.1371/journal.pone.0240362.g001

Fig 19 also provides some descriptive statistics on consumers’ perceptions and on the adoption of non-banking services and social networks by gender, age, and employment status. 91.3% of participants are frequent internet users. In terms of digital equipment, 74.8% of them reported having a laptop, 98.5% reported having a mobile phone and 46.8% reported having a tablet. In any event, the figures suggest that Spanish consumers have attained a medium-high degree of digitalization and a medium degree of financial digitalization. 79.5% of participants have an online bank account. In general, it seems that adults under the age of 45 (working or studying) are the most digitalized. It does not seem to be a gender gap in terms of financial digitalization. More significant differences emerge by employment status as working people tend to be more digitalized than the unemployed. In terms of perceptions, while most of the people perceive online and mobile banking as having a low or very low cost and safe or very safe, this percentage is smaller for those above 65 years. Finally, as expected, young people and users of social media are also more frequent adopters. 50.7% of young people (18–24 years old) have a non-bank account to make payments and 91.1% of them are active users of social networks. However, it seems that social networks such as Facebook and Twitter are seldomly used to interact or to express a complaint to the provider of financial services.

3.3 Dimensions of the digitalization process

Going digital is a much broader concept than is commonly understood. Digitalization is not a single dimensional technological expansion but a multifaceted phenomenon. While literature about the global digitalization of societies has examined several dimensions of the digitalization process [ 55 – 57 ], previous studies on the financial digitalization of consumers have mainly focused on the adoption of online channels. As the OECD has suggested, it is convenient to apply a examine a number of dimensions to explore the digital transformation of bank customers. Furthermore, prior findings in the context of online banking—a variety of mechanisms are involved in technology adoption, and the relevance of each one depends on the stage of the adoption [ 15 ]—suggest exploring more than one dimension to address issues related to digitalization. Consequently, our study assumes a broad definition of adoption that considers not only the first use of a certain service but also its scope and frequency. Fig 2 plots the main dimensions that we identified from earlier studies (see among others [ 9 , 10 , 13 , 48 , 58 – 62 ]): adoption of digital banking, diversification of use, and adoption of bank and non-bank payment instruments. For each dimension, the number of classes is equal to the number of categories in which the individuals are classified (see S3 Appendix in S1 File ).

Fig 2 plots the dimensions of bank customers’ digitalization: adoption of digital banking, diversification of use, and adoption of bank and non-bank payment instruments.

https://doi.org/10.1371/journal.pone.0240362.g002

• Adoption of Digital Banking.

Regarding the adoption of digital banking we examine 3 classes: non-users, occasional users and incipient users. Non-users are defined as those who over the course of the year have not adopted any kind of financial digitalization, including those who are not even digitalized consumers (i.e., they do not use the internet). Respondents who have become digital customers and conduct online banking activities, but not on a monthly basis, are classified as occasional users. Finally, frequent users are those who conducted online financial activities every month over the course of the year. Fig 3 shows that 59% of the survey participants are frequent users of online financial services, which is consistent with the growth of online banking in Spain officially reported by the European Digital Agenda monitoring exercises.

Fig 3 reports the number of surveyed participants for each dimension considered. The total number of surveyed participants is 3,005.

https://doi.org/10.1371/journal.pone.0240362.g003

• Diversity of Digital Use.

While the initial phase of the digital transformation of consumers involves regular online access, going digital is also related to consumers’ use of diverse digital services. Going digital therefore means conducting a number of financial activities online and not just a single online activity (e.g., just checking one’s account balance).

The factors that drive consumers’ digital diversification might be different depending on the capabilities of the electronic device used to access the service. Therefore, we differentiate between the diversification of online banking users and mobile banking users. As such, survey respondents are classified according to the variety of tasks they carry out (check account balances, pay bills, make transfers, or receive communications). Based on these factors, respondents are then sorted into four categories: no digital users, non-users of digital financial services, incipient users and diversified users.

Individuals who are outside of the digitalization process (i.e., who have no access to the internet) are classified as no digital users. In case of mobile banking, those who do not own a smartphone are also classified as no digital users. Individuals who are frequent internet users but do not conduct any financial activity online are classified as non-users of digital financial services. Incipient users are those who perform some but not all online financial activities at least once a month. Finally, those users that carry out all financial activities online at least once a month are classified as diversified users of digital financial services. Fig 3 reveals that most of the respondents are incipient users, which reflects the worth of exploring this dimension. Bank customers also appear to be customers of digital financial services, but they are still far from being considered “omni-digital” users.

• Use of Banks’ Payment Instruments.

Although debit and credit cards cannot be considered fully new electronic payment instruments, we also consider them because there has been a technological and safety evolution (i.e. contactless technology). Individuals are divided into 2 classes: non-debit (non-credit) card users and debit (credit) card users. As Fig 3 shows, there is a larger use of debit cards (78%) in comparison to credit cards (51%).

• Use of Non-Bank Payment Instruments.

While banks have traditionally offered non-cash payment instruments, some technology companies, particularly BigTech and FinTech, have begun to offer non-banking alternatives to pay bills or transfer. Since the adoption of these new means of payment provided by non-financial entities has gained ground, it is interesting to analyze how consumers adopt these alternative means of payments. A non-banking payment instrument is the one that is provided by non-banking institutions (Amazon Pay, PayPal, Google Wallet, Apple Pay, etc.). Regarding the use of these non-bank payment options, 3 classes are considered: non-digital users, non-users of non-bank payment instruments and users of non-bank payment instruments.

Consumers who do not use the internet regularly were classified as non-digital users. Consumers of online financial services who do not use non-bank means of payment were classified as non-users of non-bank payment instruments. Finally, users of non-bank payment instruments include consumers that utilize payment methods of non-bank providers. As illustrated in Fig 3 , most respondents are non-users of non-bank payment instruments despite being digitalized.

4. Methodology

The human participation in this study was voluntary. It consisted of a telephone survey conducted with all the legal and sociological guarantees by the specialised firm IMOP, as stated in the manuscript. The consent of all the survey participants was informed before conducting the questionnaire. This consent was documented as part of the recorded telephone survey which is guarded by IMOP Insights (C/ Antracita n° 7 Planta 4, 28045 –Madrid). The consent has been verified to be legal according to the Funcas’s Ethics committee and approved by the same Funcas's Ethics Committee, which also ensures that the survey has been conducted according to the principles expressed in the Declaration of Helsinki. Data were analyzed anonymously by the authors.

Most previous studies have employed discrete choice models to examine consumer preferences regarding payments and other financial services [ 6 , 14 , 63 ]. These models, derived from utility theory, are based on maximizing consumers’ utility. Other studies have used structural equations. These structural equations are useful for imputing relationships between latent variables that affect e-banking adoption [ 12 , 15 , 43 ].

However, recent studies have shown that digitalization is a challenging endeavor since there are several and complex factors driving the digitalization of people [ 20 ]. Then, this complex patterns suggests that a multidisciplinary approach is required to address digitalization [ 5 ]. In doing so, as Delen & Zolbanin [ 22 ] argue, machine learning techniques complement to the traditional research methods to address this sort of research questions. Machine learning methods are powerful tools for data mining and permit to take new insights into consumer behavior [ 21 , 23 , 24 ]. In the context of bank customers’ digitalization, machine learning would allow to reveal the complex patterns driving the digitalization process as these algorithms are able to identify complex and nonobvious patterns or knowledge hidden in a database with millions of data points.

In this sense, Bajari, Nekipelov, Ryan, and Yang [ 64 ] survey a number of methods used in demand studies to conclude that machine learning techniques are both adequate and effective for this type of analyses as they reveal complex patterns. The advantages of following a machine learning approach in complex contexts, such as consumer behavior, would explain why this machine learning is gaining ground. Miguéis, Camanho, and Borges [ 65 ] use a random forest model to find hidden patterns that may be valuable for decision-making in bank marketing. Among others, a machine learning approach is employed to estimate consumer preferences for technology products [ 26 ], to examine travel choices [ 27 ] or, more generally, to model consumer response [ 21 ]. These studies as well as other related research [ 66 – 70 ] indicate that, in a context similar to ours, machine learning algorithms provide greater accuracy (compared to other standard approaches).

Instead of ex-ante selecting a default machine learning technique, we initially consider a number of machine learning techniques that have proved their value in this field (see among others [ 28 – 32 ]. The following machine learning approach is employed to examine bank customer digitalization:

• Compare a number of machine learning methods arising from many different families and areas of knowledge to select the method that achieves the best performance in terms of accuracy.
• Employ the selected algorithm to identify the main features driving the bank customer digitalization process.
• Build a set of classification trees to illustrate the sequence of consumers’ decision-making actions.
• Use a causal machine learning technique (causal forests) to estimate causal relationships in the digitalization process.

Step 1 and 2 allow us to identify the main features driving the bank customer digitalization process based on the machine learning algorithm with better predictive performance. This way we avoid biases from ex-ante self-selecting a machine learning model. Step 3 allows us to go further in the analysis of bank customers’ digitalization. By estimating a conditional inference tree for each dimension, we may explain the decision-making process. Finally, step 4 allows us to use a causal machine learning algorithm to estimate the impact of the features with the larger predictive power on the digitalization process. Unlike prior studies, this approach allows us to examine what characteristics have a predictive power in explaining the digitalization process (step 1 and 2) but also to explore the decision-making process and the potential effect of these features (causality) on going digital (steps 3 and 4).

All the empirical analyses conducted in the paper are carried out using R software. In each and every case, the models are fed with all the variables extracted from the survey (94 variables—S2 Appendix in S1 File ), excluding the outcome. This is the a common procedure in the literature when data comes from a survey specifically designed to examine digital banking (see among others [ 17 , 71 – 73 ]). Moreover, as it has been argued in the literature, if the input variables that feed the algorithms are ex-ante filtered or chosen by the researcher, the results obtained would be biased due self-selection process.

Additionally, for those machine learning techniques that require selecting some hyperparameters (e.g. number of features for each tree in the random forest algorithm or C and gamma values in the SVM), they are not arbitrarily chosen but tuned to obtain the optimal parameter values for higher accuracy. The performance of all the machine learning methods and the logit models is computed after having optimized the hyper-parameters for each and every method. In doing so, the following R packages are employed: tune, caret, tuneRF and xgboost.

4.1 Machine learning techniques

• random forest..

Random forests are an ensemble of tree predictors in which each tree depends on the values of a random vector sampled independently and with the same distribution for all trees within the forest. Because of the law of large numbers they do not tend to overfit [ 74 ]. The algorithm follows these steps:

• A forest of many trees is grown. Each tree is grown from an independent bootstrap sample derived from the data.
• For each node of the tree, m variables are independently selected at random out of all M possible variables. Then, on the selected m variables it finds the best split.
• The algorithm grows each tree to largest extent possible.
• These steps are iterated over all trees in the ensemble, and the average vote of all the trees is reported as the random forest prediction.

• Extreme Gradient Boosting.

Gradient boosting is a machine learning technique for regression and classification problems. It came out of the idea of whether a weak learner can be modified to become better. As Valiant [ 75 ] argues, the weak learning method is used several times to get a succession of hypotheses, each one refocused on the examples that the previous ones found difficult and misclassified. Then, using a training sample (y, x) the goal of the algorithm is to obtain an estimate of the function F(x) that minimizes the expected value of a loss function over the joint distribution of all the observed values.

Among the gradient boosting methods used in practice, the Extreme Gradient Boosting, is widely used as it is an efficient implementation of the gradient boosting framework. The most important factor behind the success of the extreme gradient boosting is its scalability in all scenarios. Compared to other gradient boosting methods, the extreme gradient boosting use a more regularized model formalization to control over-fitting [ 76 ].

• K-Nearest Neighbor.

The k-nearest neighbors (k-NN) algorithm is a supervised machine learning technique employing a non-parametric method [ 77 ]. This algorithm assigns points to the data, compares them using a distance function, and assigns a classification based on the labels of the nearest points. The data point which is located at the minimum distance from the test point is assumed to belong to the same class. One of the advantages of this algorithm is that it does not derive any discriminative function from the training data. That is why the k-nearest neighbor algorithm is called a lazy learner or Instance based learning. Moreover, the k-NN algorithm is robust to data that contains a lot of noise and it is able to handle data with multiple classes.

• Support Vector Machine.

A support vector machine (SVM) is a supervised machine learning model that uses classification algorithms to solve a prediction problem for a discrete outcome using a vector of regressors, initially developed by Vapnik [ 78 ]. The algorithm constructs an optimal hyperplane that correctly classifies data points by separating the points of categories as much as possible [ 79 ]. The closest values to the classification margin are known as support vectors while the goal is to maximize the margin between the hyperplane and the support vectors [ 80 ].

Empirically, the kernel used in training the support vector machine includes the linear, radial, polynomial and sigmoid functions.

• Bayesian Networks.

Bayesian network classifiers [ 81 , 82 ] are competitive performance classifiers [ 83 ]. In this sense, a Bayesian network classifier is simply a Bayesian network applied to classification. Specifically, the prediction of the probability of some discrete (class) variable Y given some features X. Together with the well-known Naive Bayes classifier [ 84 ] more elaborate models exist taking advantage of the Bayesian network [ 85 , 86 ] such as the averaged one-dependence estimators (AODE) [ 87 ], the Chow-Liu’s algorithm for one-dependence estimators (CL-ODE) [ 81 ], the forward sequential selection and joining (FSSJ) [ 88 ], the backward sequential elimination and joining (BSEJ) [ 88 ], the Hill-climbing tree augmented naive Bayes (TAN-HC) [ 89 ] and the Hill-climbing super-parent tree augmented naive Bayes (TAN-HCSP) [ 89 ].

• Artificial Neural Networks: Extreme learning machine

Extreme learning machine (ELM) is a type of artificial neural network, called feedforward neural networks, which randomly chooses hidden nodes and analytically determines the output weights of single-hidden layer feed forward neural networks (SLFNs). The learning speed is thousands of times faster than traditional feedforward network learning algorithms like the back-propagation (BP) algorithm [ 90 ]. Moreover, compared with the conventional neural network learning algorithm it overcomes the over-fitting problem [ 91 ]. Mathematically, given a training set, an activation function and a hidden node, the algorithm follows three main steps:

• It assigns randomly input weight w i and bias b i , (i = 1,… N).
• It calculates the hidden layer output matrix
• It calculates the output weight

The activation function commonly used include the sigmoidal functions as well as the radial basis, sine, hard-limit, symmetric hard-limit, satlins, tan-sigmoid, triangular basis, rectifier linear unit and linear function.

4.2 Logit Model

Since prior literature has mainly employed discrete choice models to examine customers’ behavior, we also employ logit models to examine bank customer digitalization where Y {\ displaystyle y {*} } YY is the level of bank digitalization for each dimension of financial digitalization considered, X = ( x 1 ,…, x n ) is the set of variables and i = (1,… j) are the different categories for each dimension.

We employ an ordered logit regression for the adoption decision and the diversification of digital usage and a simple conditional logit—for the adoption of bank or non-bank payment instruments. To be consistent, the same set of variables used in the machine learning methods are employed.

4.3 Conditional inference trees

We use the characteristics and determinants with the largest discriminant power to build a decision tree for each dimension by estimating a conditional inference tree. This technique estimates a regression relationship by binary recursive partitioning in a conditional inference framework. In order to build the trees for each dimension, we follow the methodology developed by Hothorn, Hornik, & Zeileis [ 92 ] and Hothorn, Hornik, Van DeWiel, et al. [ 93 ]. The algorithm tests the global null hypothesis of independence between each of the input variables and the response and selects the input variable with the strongest association to the response. The algorithm then implements a binary split in the selected input variable and recursively repeated this process for the each of the remaining variables. The classification tree infers the sequencing of customers’ decision-making process, which helps to explain how bank customers go digital. This is particularly relevant since those classification trees do not require any linearity assumptions, which is important because many of the digitalization determinants could be nonlinearly related.

4.4 Causal machine learning

Since machine learning models are not designed to estimate causal effects, a new field of study has emerged very recently, the causal machine learning. Over the last few years, different causal machine learning algorithms have been developed, combining the advances from machine learning with the theory of causal inference [ 94 ]. The aim of these causal machine learning techniques is to complement the machine learning methods by estimating causal effects, rather than to substitute them [ 95 – 97 ]. The main advantage of causal machine learning is that it can be used after the modeling phase in order to confirm some of the relations between variables and the target/outcome. In our context, by employing a causal learning method we aim to examine the causal effect of those features with the larger predictive power on the digitalization process.

Among the recent methods developed in the causal machine learning literature, causal forest have gained relevance [ 95 – 97 ]. Knaus et al. [ 98 ] show that causal forests perform consistently well across different data generation processes and aggregation levels. The causal forest algorithm [ 96 ] is a forest-based method for treatment effect estimation that allows for a tractable asymptotic theory and valid statistical inference extending Breiman’s random forest algorithm.

Methodologically, causal forests maintain the main structure of random forests—including recursive partitioning, subsampling, and random split selection- but instead of averaging over the trees they allow to estimate heterogeneous treatment effects (causality) [ 99 ]. Then, compared to a regular decision tree, the causal tree uses a splitting rule that explicitly balances two objectives: first, finding the splits where treatment effects differ most, and second, estimating the treatment effects most accurately. In order to obtain consistent estimates of the treatment effects (the features that may have an impact on digitalization) it splits the training data into two subsamples: a splitting subsample and an estimating subsample [ 97 , 99 ]. The splitting subsample is used to perform the splits and thus grow the tree and the estimating subsample is then used to make the predictions. All observations in the estimating subsample are dropped down the previously grown tree until it falls into a terminal node. So, the prediction of the treatment effects is then given by the difference in the average outcomes between the treated and the untreated observations of the estimating subsample in the terminal nodes. Athey & Wager [ 99 ] provide a full mathematical explanation on how causal forests are built for causal inference.

Using this novel empirical methodology, we are able to examine the causal effect of those features with the larger predictive power on the digitalization process. Then, the level of digitalization is not our main interest but the impact of those features with the larger predictive power on the digitalization process. All analyses are carried out using the R package grf [ 100 ]. To run this causal algorithm, we take a conservative approach assuming that the level of digitalization of the customers can be arbitrarily correlated within a bank. Sample individuals are customers of 33 different banks. Hence, the errors are clustered at the bank-level, and we have a total of 33 clusters/banks.

5.1 Model selection

In order to select the model with the best performance, being consistent with the standard practice followed in the machine learning literature, we randomly selected 70% of the data as training data (2,104 observations) and designated the remaining data (901 observations) as test data. By doing so, we are able to determine the accuracy of the model ensuring that the algorithm is actually finding real patterns in the data and not just overfitting.

The performance of the models is compared by computing several metrics. Consistent with earlier machine learning studies (see among others [ 101 – 105 ]), we use accuracy as a measure of performance. It is defined as the number of correctly predicted data points out of all the data points. Moreover, we also compute additional standards metrics: precision (the number of correctly identified positive results divided by the number of all positive results, including those not identified correctly), recall (the number of correctly identified positive results divided by the number of all individuals that should have been identified as positive) and F1 score (the harmonic mean of the precision and recall). While recall tells us about the sensitivity of the model and precision provides information about its positive predictive value, the advantage of the F1 score is that it combines both metrics. A high F1 score is a sign of a well-performing model, even in situations where you might have highly imbalanced classes. Finally, for those dimensions with three or more classes (multi-classes) we also compute the Macro F1 score which is the averaged F1.

For the sake of brevity, Table 2 reports just the results for the best model identified per machine learning method after having optimized the hyper-parameters for each and every method. The forecasting accuracy for those cases in which several models (using several kernels and activation functions) are estimated could be found in S4 Appendix in S1 File reports. Moreover, to save space in Table 2 , we just report the precision, recall and F1 score for the class which is more frequent among the survey participants (see Fig 3 ). Overall, Table 2 shows that the random forest algorithm provides the highest level of accuracy for all the dimensions considered. Random forests accurately predict 88.41% of bank customers’ online banking adoption profile, 70.11% of the diversity of digital use of online banking, 70.01% of the diversity of digital use of mobile banking, 85% (74.89%) of debit (credit) card adoption, and 76.14% of non-bank payment instruments adoption. The second best method is the extreme gradient boosting algorithm, which also present a high percentage of accuracy.

https://doi.org/10.1371/journal.pone.0240362.t002

Regarding the F1 and macro F1 scores, the random forest model also seems to provide the highest values. For example, the macro F1 score is 91.41% for the adoption of online banking. The higher F1 and macro F1 scores of the random forest, together with the highest predicted accuracy, suggest that the random forest is the machine learning method that exhibits the best performance. Furthermore, we also observe that the performance obtained using most of the machine learning techniques, but specially the random forest, outperforms the standard logit and ordered logit models.

The higher predicted accuracy of the random forest algorithm is in line with prior studies. Bajari, Nekipelov, Ryan, & Yang [ 64 ] compare several methods and based on the out-of-sample prediction error shows that the random forest is the most accurate. Similarly, Fernández-Delgado et al. [ 102 ] evaluate 179 machine learning algorithms arising from 17 families to conclude that random forests provide the best results in terms of predicted accuracy. Consequently, since in our case the random forest is the most accurate algorithm, this algorithm is employed in order to identify the main features driving the bank customers’ digitalization process.

5.2 Validity

Finally, in order to check the stability of the accuracy of the results, we employ two cross validation methods: the k-fold cross-validation and the repeated K-fold cross-validation. In doing so, the dataset is split into 10 groups (k = 10), since this value has been shown empirically to yield test error rate estimates that suffer neither from excessively high bias nor from very high variance [ 106 , 107 ]. In case of the repeated K-fold cross-validation, the data is split into 10-folds, repeating the process five times. The results reported in Table 3 confirm the validity of the models employed.

https://doi.org/10.1371/journal.pone.0240362.t003

For replicability purposes -and given that the random forest is the selected algorithm- Table 3 reports the optimal hyperparameters of this algorithm for each dimension of financial digitalization. The Out-of-bag (OOB) error remains stable if more than a thousand trees are built. Then, since the improvement is mostly insignificant, the number of trees is set to 1,000. Moreover, since taller trees allow the model to learn very specific relationships between the features splitting the nodes and our data set [ 108 ], it is important to limit the depth of the tree in order to avoid overfitting. In doing so, we allow up to 20 nodes from the root down to the furthest leaf node.

5.3 Features of the digitalization of bank customers

Employing the random forest algorithm [ 74 ] we identify the features with the largest power in predicting bank customers’ digitalization reporting the relative statistical importance of each factor in the classification of individuals by their digital profiles (Figs 4 to 9 ). The determinants and characteristics are plotted on the y -axis ranked by their absolute level of importance while their relative importance is charted on the x -axis. The mean decrease in accuracy reflects the mean loss in accuracy when each specific variable is excluded from the regression algorithm. Therefore, the determinants and characteristics with the greater mean decrease in accuracy are the most relevant for the classification of bank customers. Additionally, the mean decrease in Gini is a measure of how each feature contributes to the homogeneity between the decision trees used in the resulting random forest. Furthermore, besides reporting the mean decrease in accuracy and the mean decrease in Gini for each variable, we employ the variable selection procedure MDAMDG proposed by Han et al. [ 109 ]. It consists of 1) running the random forest algorithm and returns the mean decrease in accuracy and the mean decrease in Gini of each variable 2) ranking every variable using the mean decrease in accuracy and the mean decrease in Gini, respectively, 3) scoring each variable 4) computing the total score of each variable 5) reordering them by the total score. While, as abovementioned, all the variables extracted for the survey (S2 Appendix in S1 File ) are used to feed the algorithm as input features, for the sake of brevity Figs 4 to 9 report only the top 20 features by their relative importance. These Figures provide the rank of the variables based on the mean decrease in accuracy, mean decrease in Gini and the total score [ 109 ].

https://doi.org/10.1371/journal.pone.0240362.g004

https://doi.org/10.1371/journal.pone.0240362.g005

https://doi.org/10.1371/journal.pone.0240362.g006

https://doi.org/10.1371/journal.pone.0240362.g007

https://doi.org/10.1371/journal.pone.0240362.g008

Figs 4 to 9 report the plots showing the relative statistical importance of each feature in the classification. The left-hand side graph shows the Top 20 features by Mean Decrease in Accuracy. The centered graph shows the Top 20 features by Mean Decrease in Gini. The righ-hand side graph shows the Top 20 features by Total Score [ 109 ].

https://doi.org/10.1371/journal.pone.0240362.g009

The machine learning algorithm reveals that the first-order factors determining the adoption of digital banking are online check balance (whether account balances are checked online), number of online bank accounts , online transfers (whether the customer has made an online bank transfer in the last three months) and consciousness (degree to whether the customer is conscious that an online access is available).

These results suggest that the relevant factors in going digital are those related to customers becoming accustomed to the online channels by checking their bank account balances or transferring money and being aware that these activities can be conducted online. Bank customers’ perceptions of security, cost, and ease of use of banking services were found to be secondary factors in going digital. As in other industries, consumers tend to go through several stages of adoption: awareness, consideration, and choice. Our results confirm the significance of awareness in the multistage process of going digital.

• Diversity of Digital Use: Online and Mobile Banking.

Figs 5 and 6 show the baseline random forest results in terms of the diversification of online and mobile banking services, respectively. The number of online bank accounts , consciousness (being aware of the possibility of having access to online services), safety of online banking (how customers perceive the level of security of online banking) and online banking communication (whether customers have used online services or e-mail as their communication method with their bank) are the features with the largest influence on diversifying the use of online banking.

Considering both the adoption and diversification of digital use, we argue that the digitalization process originates from the customers’ need to check their bank account balances and transfer money. However, being aware of the possibility of accessing financial services through online banking and the perceived safety of operating online are the main factors to diversify the use of online banking services. Furthermore, the digitalization of the communication channel between customers and banks also fosters the diversification of customers’ online activities.

Regarding the diversification of the use of mobile banking, we find that the factors with the greatest predictive power are the number of online bank accounts , safety mobile banking , consciousness and transferring money via mobile .

Overall, the algorithm reveals that online and mobile diversification are driven by common features: consciousness of the possibilities offered by digital banking, the perceived level of security of the channel used, and the number of digital bank accounts available. However, it is worth noting that transferring money was a distinct factor in determining the diversification of mobile banking. It seems that money transferring via mobile may become the gateway to other digital financial activities. This finding partially explains the importance of the irruption of FinTech companies in the payment sector compared to other financial services.

• Use of Banks’ Payment Instruments: Debit and Credit Cards.

The main factors that influence the use of debit and credit cards (see Figs 7 and 8 ) are the perceived cost , safety , acceptance and convenience of these payment instruments. Unlike the adoption and penetration of online and mobile banking, the use of debit and credit cards seems to be dominated by bank customers’ perceptions of cards’ cost, safety, and acceptance. It is interesting to see that merchants’ acceptance of debit and credit cards as payment instruments is relevant since it determines their utility, which could explain why bank customers are concerned about ensuring their acceptance before adopting them as regular payment instruments. This result suggests that the technological changes linked to cards (CVC code, EMV chips, contactless technology, multi-factor authentication) have been evolving and affect customers’ perceptions of safety and convenience.

Fig 9 illustrates that the adoption of non-bank payment methods is driven by mobile payment app (whether customers’ use of mobile apps to make payments), frequency and degree of online banking , online banking complaint ( whether customers’ use online channels to lodge a complaint with the bank) and being active on social media ( Twitter and/or Facebook user) . These findings reveal that the prior profile as digital bank customer (frequency and scope using online banking) as well as being already using payment apps determine the use of alternative payment methods. Moreover, the relevance of using online channel to complain may reveal that a certain level of dissatisfaction with the bank may lead bank customers to adopt non-bank means of payment.

Overall, while prior theories and studies have given prominence to the technological components of the service and to consumers’ perceptions to explain the digital jump (see among others [ 10 , 12 , 13 , 34 , 110 ]), our approach reveals that customers go digital first for information-based needs and, later, to undertake transactional services. Customers’ perceptions also play a role but only to explain the scope of the digitalization (being a diversified digital customer). However, customers’ perceptions (in particular, safety and cost), are particularly related to the use of bank payment methods (credit and debit cards). Moreover, the adoption of non-bank payments seems to be driven by the prior adoption and usage of online and mobile banking services.

• Robustness and stability over subsamples.

Finally, for robustness purposes we also employ the second best algorithm in terms of accuracy ( Table 2 ), the extreme gradient boosting, to identify the features with the largest predictive power. The figures in S5 Appendix in S1 File plot the most important features that predict bank customer digitalization based on this algorithm. The relative importance of each feature is computed using the contribution of the corresponding feature for each tree in the model (Gain). Overall, they show that the features with the largest predictive power according to the random forest algorithm are also identified as the most important by the extreme gradient boosting algorithm. Since both methods coincide on the main customers’ features predicting the level of digitalization, this adds robustness to the ability of machine learning methods to reveal the characteristics that drive customers’ digitalization.

Furthermore, we also aim to ensure that when feeding different data to the algorithm the predicted accuracy was stable. In doing so, we employ different subsamples based on socio-economics characteristics—gender, age, and habitat—to go through the machine learning process in order to show the robustness in terms of accuracy. Young people are those between 18 and 34, while old people are over 55 years old. The rural areas category includes people living in municipalities with less than 10,000 inhabitants while the urban category includes those living in cities with more than 200,000 inhabitants.

Fig 10 shows the accuracy across the different subsamples- based on gender, age, and habitat—that feed the model. As it could be observed the performance of the algorithm across these three subsamples remains similar to the whole performance when the algorithm is fed with the entire dataset. This result shows that the performance of the algorithm when examining the digitalization of bank customers is stable, which means, that it is not dependent on the sample subset used to feed the model. This is relevant since it reveals that the machine learning algorithm does not overfit bank customers’ digitalization for a particular profile of customers.

Fig 10 plots the predicted accuracy of the random forest algorithm across the different subsamples: based on gender (male vs female), age (young vs old) and habitat (rural vs urban areas).

https://doi.org/10.1371/journal.pone.0240362.g010

5.4 Bank customers’ digitization trees

The characteristics and determinants with the largest discriminant power are employed to estimate a conditional inference tree for each dimension. This technique estimates a regression relationship by binary recursive partitioning in a conditional inference framework. As already mentioned, these trees are built following the methodology developed by Hothorn, Hornik, & Zeileis [ 92 ] and Hothorn, Hornik, Van DeWiel, et al. [ 93 ]. In doing so, those variables with the largest relative importance based on Han et al. [ 109 ]’s total score, which accounts for mean decrease in accuracy and mean decrease in Gini, are selected (those variables are colored in a different color in Figs 4 to 9 ).

• Tree: Adoption of Digital Banking.

Fig 11 shows that although the range of services available online is wide, the adoption of online banking seems to emerge from customers checking their account balances. It is only after customers check their account balances that they move into transferring money online. Bank customers who do not perform either of these activities are classified as occasional or low frequency users (Node 5). Comparing those individuals who only check their account balances (Node 10) with those who only transfer money (Nodes 7 and 8), checking account balances appears to be more decisive. Furthermore, when customers begin to make transactions and are largely aware of the online possibilities, they become frequent users (Nodes 14 and 15).

https://doi.org/10.1371/journal.pone.0240362.g011

An overview of the random models and the classification trees suggests that the main channel by which bank customers become frequent users of online banking services is by their need to check their account balances and, subsequently, transfer money. Consciousness of the availability of online possibilities is also important for the customer to become a frequent digital bank user. Furthermore, the perceived safety of online banking services is not a primary determinant in becoming a frequent user. As we show in the next subsection, safety only becomes influential when customers consider conducting a wide range of transactions online.

• Tree: Diversity of Digital Banking Use.

Fig 12 reveals the relevance of the perceived security of online banking in influencing customers’ use of online financial services (Branch 2). Customers who do not consider online banking safe are not likely to become diversified users of online services (Nodes 14–21). Together with safety, customers’ use of digital channels for information purposes and their awareness of the range of online services are key determinants of the diversification of digital services demanded (Node 11). However, consciousness does not compensate for the perceived lack of safety. At most, being conscious make customers switch from non-users to incipient users (Nodes 17–21). Overall, the results suggest that while being a regular online banking user is driven by customers’ needs (e.g., checking account balances and transferring money) as well as by having a certain level of consciousness about the online possibilities, becoming a diversified digital user depends largely on the perceived level of safety.

https://doi.org/10.1371/journal.pone.0240362.g012

Fig 13 plots the classification tree for the diversity of digital use of mobile banking. The results suggest that the diversity of online and mobile banking use are driven by similar factors. The perceived level of safety of mobile banking is also relevant (Node 7). It is unlikely to find diversified users not transferring money with their phones even if they perceive mobile banking as not safe (Node 5).

https://doi.org/10.1371/journal.pone.0240362.g013

• Tree: Adoption of Bank Payment Instruments

Figs 14 and 15 plot the classification trees for debit and credit card adoption, respectively. Both trees demonstrate that safety and cost are the main drivers of adoption. Debit card users can be classified into users who consider debit cards safe, accepted, but not very convenient regardless of their cost (Node 11), and users who consider the method convenient, costless, and safe (Nodes 24 and 26). It can then be argued that a costless perception could compensate for a lack of perceived convenience.

https://doi.org/10.1371/journal.pone.0240362.g014

https://doi.org/10.1371/journal.pone.0240362.g015

In the case of credit cards, customers who perceive credit cards as unsafe regardless of their cost are less likely to use them (Nodes 14–19). Similar to debit cards, users who perceive credit cards as safe and relatively costless make up the majority of the credit card users (Node 12). The probability of adoption drops to 12% if the credit cards are considered costly.

• Tree: Use of Non-Bank Payment Instruments.

Fig 16 reveals that the adoption of non-bank payment methods occurs when customers are frequent and diversified digital banking users. For occasional and incipient online users, the likelihood of using non-bank payment instruments is quite small. However, as the frequency and diversity of use increases, being active on social media and making mobile payments increases the likelihood that customers would use non-bank payment channels. However, it is worth noting that frequent online users do not use non-bank payment methods if they are just incipient users (Node 23); it is necessary for customers to undertake several digital financial activities to jump into non-bank payments. Similarly, digital banking users who do not have frequent online access are not regular adopters of non-bank payment methods (Nodes 7, 16, 17, and 28).

Figs 11 to 16 plot the decision trees of bank customer digitalization by estimating a conditional inference tree using those features having the largest predictive power according to the random forest algorithm.

https://doi.org/10.1371/journal.pone.0240362.g016

• Robustness: Bayesian network.

We also estimate a Bayesian network based on the hill-climbing algorithm, using the subset of features with the largest discriminant power. Bayesian networks could be defined as graphical models of the relationships among a set of variables. These networks use Bayesian inference for probability computations with the aim to model conditional dependence, and therefore causation, by representing conditional dependence by edges in a directed graph [ 111 ]. All the graphs are shown in S6 Appendix in S1 File .

Regarding the adoption of online banking, the Bayesian network reveals that checking account balances online and making online transfers are parents of adopting digital banking. Interestingly, the network also reveals that checking account balances online is additionally a parent of making online transfers. This finding suggests that while both kind of activities play a role in the adoption of digital banking services, informational activities (checking account balance) may also foster customers to conduct transactional activities (online transfers). In a way, this result complements our finding that the adoption of digital banking services begins with information-based services (e.g., checking account balance), and is then followed by transactional services (e.g., online/mobile money transfer). Moreover, the Bayesian networks also reveal that effect of being conscious of the range of services that could be conducted online is related to the number of online bank accounts that a customer hold.

Regarding the use of cards as payment methods, it could be observed that the perceived cost of debit and credit cards is a parent of their use. In case of credit cards, the perceived safety is a parent of paying regularly with them. However, for debit cards safety is mediated by customers’ perceptions about cost and convenience. This finding would suggest that while the perceived cost has a direct relationship for both type of cards, it does not seem to be the same in case of the perceived safety. As for the adoption of non-bank payment methods, the network shows that being a diversified digital banking user has a direct relationship on paying with non-bank payment instruments. Additionally, being a Facebook user is a common parent of using non-bank payment instruments, together with being a diversified digital banking user, indicates the presence of interactions between social media and the degree of use online banking in paying with non-bank payment methods.

5.5 Causal effects on bank customers’ digitalization: Causal forests

Fig 17 shows the average treatment effect estimations–average differences in the level of digitalization—for those variables identified with the largest predictive power by the random forest. Applying this causal forest algorithm, since the estimated average treatment effects are positive and significantly different from zero, it could be argued that these features drive customers’ levels of digitalization. Then, causal forests reveal that for each of the dimensions examined those features with the largest predictive power also have a large positive effect on the digitalization process. Interestingly, the estimation of the average treatment effects also reveals that checking online balances had the largest effect on adopting online banking while making money transfers with one’s smartphone seems to be relatively more important in order to become a diversified mobile banking customer. Moreover, regarding the use of bank payment methods, we observe that the perception of safety has the largest impact on using credit cards while the perception of cost and convenience have the largest impact on paying with debit cards. This latter result was also highlighted by the Bayesian networks. Finally, regarding non-bank payment methods, the largest effects on adoption come from being a frequent and diversified digital bank customer.

Fig 17 shows the average treatment effect estimations (ATEs) computed using the causal forest algorithm. The ATEs are shown for each dimension of bamk customers’ digitalization and for those variables with the largest predictive power according to the random forest.

https://doi.org/10.1371/journal.pone.0240362.g017

These results confirm that the digitalization of bank customers is largely affected by informational (checking account balances) and transactional activities (online/mobile transfer) while the consciousness about the range of the online services available and the safety of perception have a positive impact on diversifying the use of digital channels.

6. Discussion

6.1 supply side explanations.

While the variable capturing each customer’s bank does not rank among those with the largest importance, we aim to confirm that the digitalization process is primarily driven by consumers’ characteristics and not by their bank’s characteristics. We then re-run the machine learning algorithm for different samples of consumers aggregated by their main bank characteristics to determine whether or not the predictors and decision trees obtained are qualitatively similar to those obtained in the baseline random forests regressions.

Firstly, since bank size (market power) may play a role in digitalizing customers, we re-run separate regressions for customers of large banks with the largest customer bases in Spain: Santander, BBVA, and CaixaBank. Furthermore, we also conduct a within-bank comparison. This type of analysis helps to ensure that digitalization is not mainly driven by supply-side factors since all the consumers from each subsample would have the same supply level of digitalization. In addition, since the closure of bank branches may force some bank customers to go digital, we also check whether or not bank closures drive digitalization. In doing so, separate regressions are estimated for those customers whose main bank closed at least one branch in their province.

Fig 18 reports the relative importance—measured by mean decrease in accuracy—of those variables with the largest predictive power for the adoption of online banking. The full results for the rest of the dimensions are not reported for the sake of simplicity. Checking balances, transferring money, and being conscious of online banking and the number of one’s online accounts are consistently reported as the variables with the largest predictive power across different subsamples. Hence there are not significant differences in the predictive power of the main drivers of adopting of online banking by supply-side factors (banks’ characteristics) nor by the closure of b branches. Similarly, no qualitative differences in the relative importance of the predictors and decision trees obtained are found for other dimensions of digitalization.

Fig 18 reports the relative importance—measured by mean decrease in accuracy—of those variables with the largest predictive power for the adoption of online banking by banks’ characteristics: size (large banks’ customers—Santander, BBVA, and CaixaBank—vs other banks’ customers) and branch closure (customers whose main bank closed at least one branch in their province vs customers whose main bank has not closed any branch in their province). The bottom panel shows the predicted accuracy.

https://doi.org/10.1371/journal.pone.0240362.g018

Fig 19 shows the main characteristics of the survey participants by degree of digitalization, degree of financial digitalization, perceptions on mobile and online banking and the use of non-bank services and social networks. Results are presented by gender, age, and employment situation. Each cell represents the percentage of people over the total number of people belong to this category.

https://doi.org/10.1371/journal.pone.0240362.g019

These results have interesting business implications as they suggest that the digitalization process is mainly driven by consumer characteristics. This would imply an opportunity for banks to segment customers in order to get on board and retain digital customers. Moreover, the limited impact of the closure of bank branches on digitalization suggests that the digitalization process does not emerge because customers are forced to use digital banking when there are no physical branches to reach out. It seems that customers go digital by their own needs and perceptions not because there are fewer physical branches close to where they live.

6.2 Implications, limitations, and scope for future research

Facing digital transformation successfully is among banks’ top priorities. Digital banking is likely to soon become the main channel through which customers interact with their banks. Understanding how customers face the digital jump would help banks to retain their current customers and attract more digital users by, for example, improving those functionalities related to information and transaction-based services. However, since bank customers’ digitalization seems to be explained by the needs and perceptions of consumers, bank marketing strategies should have these dimensions into account. Similarly, the results of the study will help banks understand how their customers could potentially adopt digital payment methods offered by new competitors such as BigTech and FinTech firms.

Just like any other research work, our study has certain limitations. Despite employing is a representative testing ground for research on banking digitalization, it would be ideal to know the digitalization timing of each bank customer in order to provide further insights into the temporal structure of the digitalization process. Our findings are found to be applicable to countries with deep internet penetration, a highly banked population, and a growing use of electronic banking among consumers (e.g. Germany, France, Sweden, United Kingdom, Finland, Italy, United States, Japan, Turkey or Australia). Therefore, it would be interesting to explore whether emerging economies may face the same banking digitalization process documented in this study. It should also be acknowledged that examining bank customers’ digitalization using questionnaire data may involve some biases. In any event, we use a questionnaire that follows the structure of a well-established survey, the Survey of Consumer Payment Choice (SCPC).

Despite these limitations, we believe that the results of this study are valuable for other researchers and practitioners interested in understanding how people go digital. Overall, our study confirms the need to conduct research that covers the entire digitalization process rather than focusing on a single dimension. In addition, our research finds that the application of machine learning techniques on consumer research provides more accurate results that improve the understanding of complex topics.

7. Conclusion

Modern societies are undergoing a rapid digital transformation. A sizeable part of this change is related to the demand for financial services. The use of electronic devices such as smartphones, laptops, and tablets to conduct many financial activities has risen sharply. While the banking industry is aware of this transformation, adjusting the supply side depends on related changes in demand.

Understanding the process of financial digitalization is valuable for the banking industry to design strategies that bring on board and retain digital users. It would help banks to obtain information on how they can face competition from new providers of financial services (BigTech and FinTech). Additionally, policymakers may use this knowledge to implement more efficient policies to promote financial digitalization and enhance financial inclusion and literacy. To reach this end, this paper employs a machine learning approach to reveal the patterns driving the digitalization process and to offer a multi-dimensional comprehensive picture of the process by which bank customers become digitalized. While most previous studies have discussed the determinants of certain adoption decisions, we outline the sequence of steps that customers follow to adopt digital financial services and become diversified users. Several dimensions are considered: adoption of online banking, diversification of the use of online services, and the choice of bank versus non-bank payment instruments. Our approach benefits from the advantages of machine learning techniques, including the capacity to identify complex and nonobvious patterns or knowledge hidden in a database with millions of data points. These techniques are applied to an in-depth consumer survey specifically designed for the purpose of this study. Furthermore, we run causal forest models to examine the causal relationships on the digitalization process.

The empirical results suggest that the digitalization process is originated from customers’ need to gain information about basic aspects of their banking accounts (e.g., checking their account balances), and this facilitates a transition to transactional services (e.g., transferring money). We also find that once the initial adoption has taken place, the diversification of online and mobile services adopted by the customers becomes larger when they are conscious of the range of possibilities provided by the bank and when they perceive those options as safe. Taken together, these results suggest that while customers’ perceptions are important on using digital channels, in banking the adoption is primarily driven by information-based services. Furthermore, we show that the adoption of non-bank payment instruments (e.g., PayPal and Amazon) happens when consumers are already diversified digital bank customers. Users of non-bank payment instruments seem to have previously reached a substantial degree of banking digitalization. This suggests a certain degree of complementarity between bank and non-bank digital services.

The causal machine algorithm reveals that among the information-based activities, checking online balances has the largest effect on adopting online banking. Similarly, making money transfers with a smartphone is the transactional-based activity that is relatively more important to define a diversified mobile banking customer. These results are confirmed by Bayesian networks, which also indicate that the relevance of interactions between social media and the degree of use online banking and non-bank payment methods. Importantly, we find that the digitalization process is not mainly driven by bank characteristics. We report a limited impact of the closure of bank branches on digitalization, which suggests that customers go digital by their own needs and perceptions not because there are fewer physical branches close to where they live (a diminishing role of geographic distance in banking).

These findings are relevant to better understand the digital transformation of consumers. While prior theories and studies have given prominence to the technological components of the service and to consumers’ perceptions to explain the digital jump, our machine learning approach reveals that customers go digital first for information-based needs and, later, to undertake transactional services.

Overall, the findings of the study suggest that financial providers could benefit from the digitalization phenomenon by offering services that better match customers’ needs. In this sense, segmenting customers using similar techniques and data, would make possible to offer them more personalized digital services. Moreover, linking payments experiences to social media interactions could also be used to foster the adoption of digital payments. Finally, our findings could be used by policymakers to improve the communication and social awareness of the range of online services available, as part of the policies and official strategies to promote financial digitalization.

Supporting information

https://doi.org/10.1371/journal.pone.0240362.s001

Acknowledgments

The authors are very grateful for comments from Sumit Agarwal, Charles Kahn, Thorsten Beck, Dalida Kadyrzhanova, Glenn Harrison, Rohan Ganduri, Meryem Duygun, Ilaf Elard, Francisco Rivadeneyra, Michael King, Stefan Güldenberg, John Holland, Robert E. Wright, Kamlesh Kumar, and Paola Bognini as well as conference participants at the 15 th International Conference WEAI-IBEFA in Tokyo 2019; the EURAM 2019 Conference in Lisbon; the IFABS 2018 Chile Conference; the 2018 Workshop on Credit Card Lending and Payments, Developments, and Innovations at the Federal Reserve Bank of Philadelphia; the 2018 Conference on Financial Stability Implications of New Technology at the Federal Reserve Bank of Atlanta; the 3 rd International Workshop on the Internet for Financial Collective Awareness & Intelligence at Glasgow University; the 2018 International Workshop on Financial System Architecture and Stability (IWFSAS) at Cass Business School; and the 1 st Banca March Workshop on Contemporary Issues in Banking.

• 1. OECD (2017), OECD Digital Economy Outlook 2017, OECD Publishing, Paris.
• View Article
• Google Scholar
• PubMed/NCBI
• 24. Witten I. H., Frank E., & Hall M. A. (2019). Data Mining: Practical Machine Learning Tools and Techniques (Third Edition). In Morgan Kaufmann (The Morgan, Vol. 53, Issue 9). Morgan Kaufmann.
• 25. Bishop C.M. (2006). Pattern Recognition and Machine Learning, Information Science and Statistics (Vol.53). Springer-Verlag New York.
• 33. Federal Reserve Bank of Atlanta. (n.d.). https://www.frbatlanta.org/banking-and-payments/consumer-payments/survey-of-consumer-payment-choice . Survey of Consumer Payment Choice.
• 35. Fishbein M., & Ajzen I. (1975). Belief, attitude, intention and behavior: An introduction to theory and research. In Reading, MA: Addison-Wesley.
• 38. Rogers E. M. (1962). Diffusion of innovations. In New York, Free Press of Glencoe.
• 75. Valiant L. (2013). Probably Approximately Correct : Nature’s Algorithms for Learning and Prospering in a Complex World . Basic Books, Inc.
• 76. Chen, T., & Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting System. ArXiv:1603.02754v3 , 1–13.
• 78. Vapnik V. N. (1995). The Nature of Statistical Learning Theory . Springer-Verlag.
• 85. Koller D., & Friedman N. (2009). Probabilistic Graphical models . Principles and Techniques . The MIT Press.
• 86. Pearl J. (1998). Probabilistic Reasoning in Intelligent Systems . Morgan Kaufmann.
• 108. Shu X. (2020). Knowledge Discovery in the Social Sciences: A Data Mining Approach. In University of California Press.
• 109. Han, H., Guo, X., & Yu, H. (2016). Variable selection using Mean Decrease Accuracy and Mean Decrease Gini based on Random Forest. Proceedings of the IEEE International Conference on Software Engineering and Service Sciences, ICSESS , 219–224.
• 111. Holmes D. E., & Jain L. C. (2008). Innovation in Bayesian Networks . Theory and Applications . Springer.

• Browse All Articles
• Newsletter Sign-Up

BanksandBanking →

No results found in working knowledge.

• Were any results found in one of the other content buckets on the left?
• Try removing some search filters.
• Use different search filters.

After 40 Years, How Representative Are Labor Market Outcomes in the NLSY79?

In 1979, the National Longitudinal Study of Youth 1979 (NLSY79) began following a group of US residents born between 1957 and 1964. It has continued to re-interview these same individuals for more than four decades. Despite this long sampling period, attrition remains modest. This paper shows that after 40 years of data collection, the remaining NLYS79 sample continues to be broadly representative of their national cohorts with regard to key labor market outcomes. For NLSY79 age cohorts, life-cycle profiles of employment, hours worked, and earnings are comparable to those in the Current Population Survey. Moreover, average lifetime earnings over the age range 25 to 55 closely align with the same measure in Social Security Administration data. Our results suggest that the NLSY79 can continue to provide useful data for economists and other social scientists studying life-cycle and lifetime labor market outcomes, including earnings inequality.

We thank Kevin Bloodworth II, Elizabeth Harding, and Siyu Shi for research assistance. The views in this paper are those of the authors and do not necessarily reflect the views of the Federal Reserve Bank of St. Louis, the Federal Reserve System, or of the National Bureau of Economic Research.

Richard Rogerson acknowledges financial support in excess of $10,000 over the last three years from the Federal Reserve Bank of Atlanta, the Federal Reserve Bank of Minneapolis and the World Bank.

MARC RIS BibTeΧ

Download Citation Data

More from NBER

In addition to working papers , the NBER disseminates affiliates’ latest findings through a range of free periodicals — the NBER Reporter , the NBER Digest , the Bulletin on Retirement and Disability , the Bulletin on Health , and the Bulletin on Entrepreneurship  — as well as online conference reports , video lectures , and interviews .

International Journal of Bank Marketing: Volume 40 Issue 7

Table of contents, affective and cognitive factors that hinder the banking relationships of economically vulnerable consumers.

The aim of this paper is to explore the affective and cognitive factors that condition banking relationships for economically vulnerable consumers and how these factors contribute…

Reputation and its consequences in Fintech services: the case of mobile banking

Recently, traditional financial institutions are facing strong competition from disruptive innovators (Fintech firms) forcing them to increasingly invest in new IT solutions to…

Antecedents of word-of-mouth communication in the life insurance industry

This study explored the antecedents and mediators of word-of-mouth (WOM) behavior in the life insurance industry and investigated how to increase policyholders' positive WOM…

The psychological antecedents of personal financial management behavior: a meta-analysis

The intent of this study is to aggregate, in a measurable form, the results of previous studies on the association between personal financial management behavior (PFMB) and six…

Do emotions, desires and habits influence mutual fund investing? A study using the model of goal-directed behavior

The purpose of this paper is to determine whether affective factors such as goal desires, positive anticipated emotions, anticipated regret and non-volitional actions like habits…

The antecedents of m-banking usage under capital controls in Greece – a mixed methods approach

Capital controls restrict cash withdrawals and international transfers, among other restrictions. The purpose of the study is to explore how capital controls have influenced…

Mobile banking service quality and customer value co-creation intention: a moderated mediated model

The objective of this paper is to explore the possible impact of mobile banking service quality (MBSQ) on customer value co-creation intention (CVCI).

Ethical reputation and retail bank selection: a sequential exploratory mixed-methods study in an emerging economy

In this study, the authors examine how a retail bank's positive, neutral, and negative prior ethical reputations influence customers' perceptions and attitudes, leading to their…

Effects of teaser rates on new credit card customers’ spending and borrowing: an empirical analysis

This study examines the effect of credit card teaser rates on consumer indebtedness and the revenue generated by new customers.

A systematic and bibliometric review of the financial well-being: advancements in the current status and future research agenda

Achieving financial well-being is essential for individuals, families and countries as it leads to life satisfaction and happiness. This study synthesizes and identifies financial…

The role of the hedonistic and utilitarian quality dimensions in enhancing user satisfaction in mobile banking

The purpose of this study is to examine the components of the utilitarian and hedonistic dimensions of the quality of mobile banking services, their impact on customer…

Towards a better understanding of the concept of revenge against banks

The purpose of this research is to investigate the concept of revenge against banks.

Why do you look familiar? - Exploring the role of employee-customer identification in adapting the service offering

This study aims to explore employee-customer identification and its consequences in the banking, financial service and insurance (BFSI) sector. We also look at the mediating role…

Financial capability: a systematic conceptual review, extension and synthesis

This study aims to examine the literature on consumer financial capability. By analyzing the research trends, theories, definitions and themes, the literature on financial…

The role of financial literacy for financial resilience in middle-age and older adulthood

Preserving sufficient financial assets is crucial for maintaining the standard of living. The lack of adequate financial cushion can translate into financial hardship at any age…

Online date, start – end:

Copyright holder:, open access:.

• Dr Hooman Estelami

Further Information

• About the journal (opens new window)
• Purchase information (opens new window)
• Editorial team (opens new window)
• Write for this journal (opens new window)

We’re listening — tell us what you think

Something didn’t work….

Report bugs here

All feedback is valuable

Please share your general feedback

Join us on our journey

Platform update page.

Visit emeraldpublishing.com/platformupdate to discover the latest news and updates

Questions & More Information

Answers to the most commonly asked questions here

Our use of cookies

We use necessary cookies to make our site work (for example, to manage your session). We’d also like to use some non-essential cookies (including third-party cookies) to help us improve the site. By clicking ‘Accept recommended settings’ on this banner, you accept our use of optional cookies.

Necessary cookies

Necessary cookies enable core functionality on our website such as security, network management, and accessibility. You may disable these by changing your browser settings, but this may affect how the website functions.

Analytics cookies

We use analytics cookies so we can keep track of the number of visitors to various parts of the site and understand how our website is used. For more information on how these cookies work please see our Cookie policy .

An approach to cleaning MiFID II corporate bond transaction reports

Staff working paper no. 1,071.

By Simon Jurkatis

Since 2018, EU and UK financial markets regulators have been in receipt of data on transactions in debt instruments, such as corporate bonds, reported under the Markets in Financial Instrument Regulation. The data gives regulators a more detailed and broader view of trading in these instruments than previously. Reports submitted under this framework, however, come with a number of unique challenges that require careful consideration. Among those challenges are that reports are not submitted in a completely standardised way, that prices and quantities can be reported in different units, and that reports may be submitted by both counterparties of a transaction. This paper describes an approach for handling these issues for transaction reports on corporate bonds, with the aim of helping to enhance the data quality and supporting robust research into this market.

Sign up for latest updates

Other staff working papers, the role of finance for export dynamics:..., the role of finance for export dynamics: evidence from the uk, monetary policy consequences of financial..., monetary policy consequences of financial stability interventions: assessing the uk ldi crisis and the..., central bank profit distribution and recapitalisation, an unconventional fx tail risk story.

Back to top

Home Toggle navigation FR Toggle Search Search the site Search About us About us Head office Regional offices History Archives Background materials Photos and videos Accessibility Contact us Corporate governance Board of Directors Governing Council and Senior Management Governance documents Educational resources The Economy, Plain and Simple Explainers Financial education resources Careers Take a central role at the Bank of Canada with our current opportunities and scholarships.

The Neutral Interest Rate: Past, Present and Future

The decline in safe real interest rates over the past three decades has reignited discussions on the neutral real interest rate, known as R*. We review insights from the literature on R*, addressing its determinants and estimation methods, as well as the factors influencing its decline and its future trajectory. While there is a consensus that R* has declined, alternative estimation approaches can yield substantially different point estimates over time. The estimated neutral range is large and uncertain, especially in real-time and when comparing estimates based on macroeconomic data with those inferred from financial data. Evidence suggests that factors such as increased longevity, declining fertility rates and scarcity of safe assets, as well as income inequality, contribute to lowering R*. Existing evidence also suggests the COVID-19 pandemic did not substantially impact R*. Going forward, there is an upside risk that some pre-existing trends might weaken or reverse.

DOI: https://doi.org/10.34989/sdp-2024-3

We use cookies to help us keep improving this website.

• Awards & recognition
• Executive biographies
• Bank of America Institute
• Media contacts
• Media content library
• Report center

Ten global investment banks and market making firms join BGC in the creation of FMX to launch premier U.S. Treasury and U.S. Interest Rate Futures trading marketplace

April 25, 2024 at 10:05 AM Eastern

FMX closes transaction with strategic investors at a $667 million post-money equity valuation

NEW YORK, NY –   BGC Group, Inc. (Nasdaq: BGC) today announced that Bank of America, Barclays, Citadel Securities, Citi, Goldman Sachs, J.P. Morgan, Jump Trading Group, Morgan Stanley, Tower Research Capital, and Wells Fargo have become minority equity owners of FMX, with a post-money equity valuation of $667 million. FMX combines BGC’s U.S. cash treasuries platform with its spot foreign exchange platform and U.S. interest rate futures exchange, and will leverage BGC’s proven low latency trading infrastructure and global distribution to further support liquidity in the interest rate futures market.  

“We have brought together ten of the most important global investment banks and market making firms to create a premier trading venue for the interest rate markets,” said Howard W. Lutnick, Chairman and CEO of BGC Group and Chairman of FMX . “We offered ownership to this incredible investment group knowing the enormous value they bring to FMX, which will benefit all market participants.”

FMX Futures, which received CFTC approval in January , is expected to launch in September 2024. FMX’s cash U.S. Treasury platform, FMX UST (formerly known as Fenics UST), has grown its Central Limit Order Book market share each sequential quarter. FMX UST ended the first quarter 2024 at 28%, up from 26% in the fourth quarter of 2023.  1

“With support from these leading fnancial firms, we believe FMX will become a rapidly growing futures platform and create important efficiencies for our shared clients,” said Lou Scotto, CEO of FMX . “With our clearing partner, LCH, the largest clearer of interest rate swaps in the world, clients will receive significant portfolio-margining capabilities, creating competitive advantages across U.S. interest rate markets.”   2

“LCH is excited to partner with FMX to deliver product innovation and margin savings, which will enhance the competitiveness of U.S. derivatives markets for its members,” said Isabelle Girolami, CEO of LCH Ltd.

"FMX's unique protocols provide a fresh competitive edge across rates, FX, and futures markets,” said Geoff Weber, Head of G10 Rates Flow Trading at Citi . “The impressive growth in market share that FMX has experienced recently enhances market liquidity and positions FMX as a potential catalyst for increased competition, particularly within the futures market. This innovation not only promises to elevate market dynamics but also aims to lower costs for all market participants, signaling a forward-looking shift."

“FMX is going to drive innovation and competition across the rates, FX and futures markets,” said Kristen Macleod, Head of Americas Macro Distribution and Co-Head of Global FX Distribution at Barclays . “As a key investor, Barclays looks forward to delivering the benefits of our investment to our clients through improved execution and competitive fees.”

Please find additional details about the FMX transaction at www.ir.bgcg.com . BGC will also provide additional information about the FMX transaction on its first quarter 2024 earnings call, scheduled for 10:00 a.m. ET on Tuesday, April 30, 2024.

About BGC Group, Inc.

BGC Group, Inc. (Nasdaq: BGC) is a leading global marketplace, data, and financial technology services company for a broad range of products, including fixed income, foreign exchange, energy, commodities, shipping, equities, and now includes the FMX Futures Exchange. BGC’s clients are many of the world’s largest banks, broker-dealers, investment banks, trading firms, hedge funds, governments, corporations, and investment firms.

BGC and leading global investment banks and market making firms have partnered to create FMX, part of the BGC Group of companies, which includes a U.S. interest rate futures exchange, spot foreign exchange platform and the world’s fastest growing U.S. cash treasuries platform.

For more information about BGC, please visit www.bgcg.com .

Discussion of Forward-Looking Statements about BGC

Statements in this document regarding BGC that are not historical facts are “forward-looking statements” that involve risks and uncertainties, which could cause actual results to differ from those contained in the forward-looking statements. These include statements about the Company’s business, results, financial position, liquidity and outlook, which may constitute forward-looking statements and are subject to the risk that the actual impact may differ, possibly materially, from what is currently expected. Except as required by law, BGC undertakes no obligation to update any forward-looking statements. For a discussion of additional risks and uncertainties, which could cause actual results to differ from those contained in the forward-looking statements, see BGC’s Securities and Exchange Commission ("SEC") filings, including, but not limited to, the risk factors and Special Note on Forward-Looking Information set forth in these filings and any updates to such risk factors and Special Note on Forward-Looking Information contained in subsequent reports on Form 10-K, Form 10-Q or Form 8-K.

Reporters may contact:

Erica Chase [email protected] +1 212-610-2419

Sofia Rehman [email protected]

Rekha Jogia-Soni [email protected]

Investors may contact:

Jason Chryssicas +1 212-610-2426

1 Central Limit Order Book (CLOB) is a mechanism financial exchanges use to facilitate trading between buyers and sellers in financial markets. Source: Coalition Greenwich.

2 Source: Clarus.

• Global Investment Banking and Global Markets

Get Bank of America news updates, including dividend announcements and other important information to your inbox.

What would you like the power to do.

Tell us about your experience

Help Center

• Location finder

Our company

• Responsible growth
• Business practices
• What we offer
• Modern Slavery Act Statement (PDF)

Making an impact

• Local Impact
• Racial equality & economic opportunity
• Driving economic & social progress
• Environmental sustainability
• Find resources

Working here

• Being a great place to work
• Quarterly earnings
• Events & presentations
• Regulatory & other filings
• Fixed income
• Shareholder info
• Annual reports & proxy
• Press releases
• Journalist resources

Investing in securities involves risks, and there is always the potential of losing money when you invest in securities.

This material does not take into account your particular investment objectives, financial situations or needs and is not intended as a recommendation, offer or solicitation for the purchase or sale of any security, financial instrument, or strategy. Before acting on any information in this material, you should consider whether it is suitable for your particular circumstances and, if necessary, seek professional advice. Any opinions expressed herein are given in good faith, are subject to change without notice, and are only correct as of the stated date of their issue.

Content contained herein may have been produced by an outside party that is not affiliated with Bank of America or any of its affiliates (Bank of America). Opinions or ideas expressed are not necessarily those of Bank of America nor do they reflect their views or endorsement. These materials are for informational purposes only. Bank of America does not assume liability for any loss or damage resulting from anyone's reliance on the information provided. Certain links may direct you away from Bank of America to an unaffiliated site. Bank of America has not been involved in the preparation of the content supplied at the unaffiliated sites and does not guarantee or assume any responsibility for its content. When you visit these sites, you are agreeing to all of their terms of use, including their privacy and security policies.

Merrill Lynch, Pierce, Fenner & Smith Incorporated (also referred to as “ MLPF&S ” or “Merrill”) makes available certain investment products sponsored, managed, distributed or provided by companies that are affiliates of Bank of America Corporation (“ BofA Corp.”). MLPF&S is a registered broker-dealer, registered investment adviser, Member  SIPC  and a wholly owned subsidiary of BofA Corp.

Trust and fiduciary services are provided by Bank of America Private Bank, a division of Bank of America, N.A., Member FDIC , and a wholly-owned subsidiary of Bank of America Corporation (“ BofA Corp.”). Insurance and annuity products are offered through Merrill Lynch Life Agency Inc. (“ MLLA ”), a licensed insurance agency and wholly-owned subsidiary of BofA Corp.

Investment products offered through MLPF&S and insurance and annuity products offered through MLLA :

Bank of America Private Bank is a division of Bank of America, N.A., Member FDIC and a wholly owned subsidiary of Bank of America Corporation ( BofA Corp.).

"Bank of America" is the marketing name for the global banking and global markets business of Bank of America Corporation. Lending, derivatives, and other commercial banking activities are performed globally by banking affiliates of Bank of America Corporation, including Bank of America, N.A., Member FDIC. Securities, strategic advisory, and other investment banking activities are performed globally by investment banking affiliates of Bank of America Corporation ("Investment Banking Affiliates"), including, in the United States, BofA Securities, Inc. and Merrill Lynch, Pierce, Fenner & Smith Incorporated, both of which are registered broker-dealers and Members of  SIPC , and, in other jurisdictions, by locally registered entities. BofA Securities, Inc. is registered as a futures commission merchant with the CFTC and is a member of the NFA.

Company goals are aspirational and not guarantees or promises that all goals will be met. Statistics and metrics included in our ESG documents are estimates and may be based on assumptions or developing standards.

• Review Cookie Categories
• Accessible Banking
• Social community guidelines
• Advertising Practices

Bank of America, N.A. Member FDIC.  Equal Housing Lender © 2024 Bank of America Corporation. All rights reserved.

Connect with us

• Site Search
• IT aliana versione
• Show the navigation menu

main sections of the site

local navigation

• you are here: Home

New releases of Occasional Papers and Newsletter of the economic research of the Bank of Italy - 26 April 2024

26 April 2024

Today the Bank of Italy has released three new Occasional Papers (nn. 846-848) and the No. 74 of the Newsletter on the economic research in Bank of Italy.

• Newsletter on the economic research No. 74 - April 2024 pdf 529.6 KB
• No. 848 - Increasing macroprudential space in Italy by activating a systemic risk buffer
• No. 847 - Climate-related risks for Italy: an analysis based on the latest NGFS scenarios
• No. 846 - Linking macro- and microdata to produce distributional accounts for non-financial corporations

you are here: Home Media News

back : News

BANK OF ITALY

• Accessing online services using SPID
• Transparency
• Working for the Bank of Italy
• Financial Intelligence Unit
• Banking and Financial Ombudsman
• Economics for everyone
• MUDEM - Money Museum
• Central bank websites
• Banking supervision portal

Legal Information

• Public Key Infrastructure
• Privacy Policy
• Cookie Preferences
• E-mail Alert
• © Bank of Italy
• VAT NO. 00950501007
• Fiscal Code 00997670583

Informativa sui cookie: La Banca d’Italia utilizza cookie tecnici e di terze parti per il funzionamento del sito: per maggiori informazioni e per sapere come abilitarli in modo selettivo leggere l'informativa sulla privacy . Utilizzare uno dei seguenti pulsanti per esprimere la propria preferenza. Se si clicca su “Chiudi/Rifiuta tutti i cookie” saranno attivati i soli cookie tecnici interni al sito.

About this site's cookies: The Bank of Italy uses technical and third-party cookies for the operation of the site: for more information and to know how to selectively enable cookies, read the Privacy Policy . Click on one of the buttons below to express your preference. If you click "Close/Reject all cookies", only the internal technical cookies of the site will be activated.

Research of Integrated Marketing Communications in the Automotive Market

• Conference paper
• First Online: 18 March 2022
• Cite this conference paper

• Olga Pitko   ORCID: orcid.org/0000-0001-6063-8514 11  

Part of the book series: Lecture Notes in Networks and Systems ((LNNS,volume 403))

Included in the following conference series:

• International Scientific Siberian Transport Forum

727 Accesses

1 Citations

The article studies the integrated marketing communications in the automotive market using the example of OOO “Bogemia Novosibirsk” (“Skoda” official dealer). The analysis of the communication channels used by the car dealer made it possible to come to the conclusion about the use of traditional promotion tools for certain segments of the target audience in the individual information offer formation. The marketing communications program for the female target audience was developed in the form of an algorithm for the communication stages. The psychological profile of female consumers was compiled on the basis of a value system study that determines the choice of passenger cars by these buyers. The result of the analysis allowed us to conclude that integrated marketing communications are fundamentally different for the female audience segments - single and married. The main methods of attracting a female consumer in favor of a particular car dealership and shaping her reaction are provided by advertising, PR (including sponsorship), sales promotion and personal sales, which ensure synergistic effectiveness of actions to position the company in the passenger car market, and, accordingly, to draw attention to her. Regarding ways of promoting sales for the target segment of women, a set of possible incentives for the car dealership, focused on immediate and delayed effects, has been developed. The complex of the presented activities is associated with their differentiated significance for the target audience segments identified during the analysis. The named communications developed in the context of relationship marketing at the stage of female consumer behavior after the purchase are becoming especially relevant.

This is a preview of subscription content, log in via an institution to check access.

Access this chapter

• Available as PDF
• Read on any device
• Instant download
• Own it forever
• Available as EPUB and PDF
• Compact, lightweight edition
• Dispatched in 3 to 5 business days
• Free shipping worldwide - see info

Tax calculation will be finalised at checkout

Purchases are for personal use only

Institutional subscriptions

Pitko, O.A.: Internet promotion of a commercial organization in the consumer market. Econ. Entrepreneurship 12 (125), 823–827 (2020). https://doi.org/10.34925/EIP.2021.125.12.163

Ali, N.A.M., Gafar, M.H.A., Akbar, J.: Enhancing promotional strategies within automotive companies in Malaysia. Procedia Econ. Finance 7 , 158–163 (2013). https://doi.org/10.1016/S2212-5671(13)00230-X

Article   Google Scholar  

Dmitrijeva, K., Batraga, A.: Barriers to integrated marketing communications: the case of Latvia (small markets). Procedia Soc. Behav. Sci. 58 , 1018–1026 (2012). https://doi.org/10.1016/j.sbspro.2012.09.1082

Reshetko, N., Safronova, A., Vakulenko, S., Kurenkov, P., Sokolova, A.: Quality assessment of management decisions in the system of marketing and public relations of a transport enterprise. Transp. Res. Procedia 54 , 380–387 (2021). https://doi.org/10.1016/j.trpro.2021.02.087

Durugbo, C., Erkoyuncu, J.: Managing integrated information flow for industrial service partnerships: a case study of aerospace firms. Procedia CIRP 16 , 338–343 (2014). https://doi.org/10.1016/j.procir.2014.01.032

Fonseca, L., Fernandes, J., Delgado, C.: QFD as a tool to improve negotiation process, product quality, and market success, in an automotive industry battery components supplier. Procedia Manuf. 51 , 1403–1409 (2020). https://doi.org/10.1016/j.promfg.2020.10.195

Najafi-Tavani, S., Sharifi, H., Najafi-Tavani, Z.: Market orientation, marketing capability, and new product performance: the moderating role of absorptive capacity. J. Bus. Res. 69 (11), 5059–5064 (2016). https://doi.org/10.1016/j.jbusres.2016.04.080

Bellou, F.: Cultural awareness in peace operations: effective marketing or strategic communications. Procedia Soc. Behav. Sci. 148 , 579–587 (2014). https://doi.org/10.1016/j.sbspro.2014.07.083

Classen, M., Friedli, T.: Value-based marketing and sales of industrial services: a systematic literature review in the age of digital technologies. Procedia CIRP 83 , 1–7 (2019). https://doi.org/10.1016/j.procir.2019.02.141

Abishovna, B.: The principle of effective marketing management. Procedia Soc. Behav. Sci. 109 , 1322–1325 (2014). https://doi.org/10.1016/j.sbspro.2013.12.632

Dieguez, T., Ferreira, L.P., Silva, F.J.G., Tjahjono, B.: Open innovation and sustainable development through industry-academia collaboration: a case study of automotive sector. Procedia Manuf. 51 , 1773–1778 (2020). https://doi.org/10.1016/j.promfg.2020.10.246

Nadzri, W.N.M., Musa, R., Muda, M., Hassan, F.: The antecedents of brand experience within the national automotive industry. Procedia Econ. Finance 37 , 317–323 (2016). https://doi.org/10.1016/S2212-5671(16)30131-9

Golfetto, F., Gibbert, M.: Marketing competencies and the sources of customer value in business markets. Ind. Mark. Manage. 35 (8), 904–912 (2016). https://doi.org/10.1016/j.indmarman.2006.06.007

Ashjaei, M., Bello, L., Daneshtalab, M., Saponara, G.S.P., Mubeen, S.: Time-sensitive networking in automotive embedded systems: state of the art and research opportunities. J. Syst. Archit. 117 , 102137 (2021). https://doi.org/10.1016/j.sysarc.2021.102137

Marco, A., Mangano, G., Magistris, P.: Evaluation of project management practices in the automotive original equipment manufacturers. Procedia Comput. Sci. 181 (2021). https://doi.org/10.1016/j.procs.2021.01.151

Download references

Author information

Authors and affiliations.

Siberian State University of Water Transport, Schetinkina Street, 33, 630099, Novosibirsk, Russia

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Olga Pitko .

Editor information

Editors and affiliations.

Rector of the Siberian Transport University, Siberian Transport University, Novosibirsk, Russia

Aleksey Manakov

Far Eastern State Transport University, Ulitsa, Khabarovsk Territory, Russia

Arkadii Edigarian

Rights and permissions

Reprints and permissions

Copyright information

© 2022 The Author(s), under exclusive license to Springer Nature Switzerland AG

About this paper

Cite this paper.

Pitko, O. (2022). Research of Integrated Marketing Communications in the Automotive Market. In: Manakov, A., Edigarian, A. (eds) International Scientific Siberian Transport Forum TransSiberia - 2021. TransSiberia 2021. Lecture Notes in Networks and Systems, vol 403. Springer, Cham. https://doi.org/10.1007/978-3-030-96383-5_77

Download citation

DOI : https://doi.org/10.1007/978-3-030-96383-5_77

Published : 18 March 2022

Publisher Name : Springer, Cham

Print ISBN : 978-3-030-96382-8

Online ISBN : 978-3-030-96383-5

eBook Packages : Engineering Engineering (R0)

Share this paper

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Publish with us

Policies and ethics

• Find a journal
• Track your research