Modelling and Evaluating Financial Sustainability using Data Envelopment Analysis and Machine Learning Methods

Shi, Yu

Etd

Modelling and Evaluating Financial Sustainability using Data Envelopment Analysis and Machine Learning Methods

Public Deposited

This dissertation aims to build the methodological framework for quantifying, evaluating, and predicting financial sustainability in the American commercial banking sector at a firm level. Financial sustainability is an organization’s ability to initiate, grow and maintain financial capacity continuously over time. Financial sustainability is comprised of various components, including accessibility to funds and other financial resources, ability to maintain the firm’s operations regeneratively, and effectively converting earnings into profitability. While extant studies that assess financial sustainability primarily focus on microfinance institutions (MFIs), which provide services to the financially marginalized population, these studies only cover a partial analysis on financial sustainability in the banking industry, as MFIs mainly target the economically disadvantaged population rather than the general public. In addition, MFIs are more prevalent in developing countries compared with those in the United States, where the role of MFIs has long been substituted by regular commercial banks that offer accessible and affordable credit or savings plans. Moreover, financial sustainability is treated as a special terminology in MFI studies, and these studies typically use bank efficiency as a proxy for financial sustainability. However, in reality, financial sustainability is a much more complex multifactor concept. The modelling and evaluation of financial sustainability in the commercial banking sector is much different from the evaluation of financial sustainability in MFIs only. A composite index integrating multiple aspects of the banks’ operations needs to be constructed, and given the multi-dimensionality of financial sustainability, the composite index should contain more than just one stage. As mentioned above, financial sustainability is a firm’s ability to maintain profitable operations in both the short run and the long run. To precisely capture and formulate financial sustainability, multiple operational aspects need to be considered and incorporated. In a typical commercial banking process, there are operations related to, for instance, attracting depositors, helping customers make deposits, using deposits to make loans and collect interests, and generating income based on the available assets and equity. Ultimately, both the financial human capital consumed by the banking process and the levels and types of outputs produced by the banking process will be pivotal to measuring how well the bank translates resources into value. This commercial banking process resembles a standard production process, where inputs are consumed, and outputs are produced. As a result, we adopt data envelopment analysis (DEA) as the primary methodology, due to its ability to evaluate and compare production units. DEA is a non-parametric production frontier analysis that can be used to benchmark production units, given their input and output levels and categories. With the help of DEA, we can treat financial sustainability as a type of performance that is tangible, convertible, quantifiable, and measurable. Nonetheless, although DEA can be used to model financial sustainability, it may be insufficient in terms of the interpretation and explanation of model results. For instance, investors, shareholders, and customers may be interested in knowing the actual factors that affect financial sustainability. To unravel the impacts of these contextual variables, we adopt machine learning methods. In essence, this dissertation first utilizes a data analytical approach to predefine and benchmark financial sustainability as a type of organizational performance. Specifically, DEA is employed due to its embodiment of an input- to-output structure inherent in all production processes. Performance metrics, including resources consumed and outputs produced, are selected manually based on empirical studies to build a composite sustainability index (CSI). A simple, one-stage CSI is first built as a preliminary experiment to test whether building such a framework is feasible. Then, because the evaluation of financial sustainability is multi-dimensional, encompassing various and even conflicting performance metrics, a more complex multi-stage network structure is developed. Then, different machine learning methods are explored, and several appropriate methods are utilized to construct methodological frameworks that predict financial sustainability in the banks’ deposit related operations, loan related operations, as well as profitability conversion. These methods include classification and feature selection using random forests, support vector machines (SVM) and logistic regression. We choose diverse methods, including an ensemble approach, a nonlinear approach, and a linear approach to examine whether our model results are consistent. Furthermore, other techniques, such as synthetic minority oversampling technique (SMOTE) and Shapley Additive Explanations (SHAP), are used to overcome problems that arise during our high- dimensional, class-imbalanced modelling. These methods altogether facilitate the interpretation and extrapolation of financial sustainability evaluation results in hopes of providing meaningful insights into the banking industry. In the end, we also corroborate the effectiveness and accuracy of these machine learning models. Overall, it is corroborated that DEA in combination with random forest classification models and SHAP analysis is the only framework that is sufficient in defining, quantifying, and accurately predicting financial sustainability in American financial institutions.

Creator