Net income is one of the most important financial indicators used to judge banks’ financial performance. It is the net returns received by the banking institutions after operating costs and various expenses are deducted; such a measure suggests a bank’s ability to use its financial resources efficiently to produce value for shareholders and customers. The greater the net income is, the more efficient (financial and operational) the management functions are, thus the higher the expansion, investment and risks the bank carries. While the banking sector is extremely sensitive (i.e. a huge reason for the importance of net income), the banking industry itself is confronted with different pressures that affect its stability and the bottom line. Foremost in those challenges are escalating operational costs caused by growing digital services and building technological infrastructure and fluctuating interest rates, which translate into decreasing interest income and expenses. Asset size and management dynamics can also make it challenging to predict financial performance in an accurate manner. Most previous research has utilised classical statistical modeling such as linear as well as multiple regression to study and forecast the source for determinants of the net income. But such models frequently do not accommodate large sample sizes and complex data and generally assume static linear relationships among the variables, which do not capture the reality of the banking industry. Thus, these traditional methods severely limit prediction accuracy and constrain decision makers to develop a good, data-driven strategy to the limit. Here lies the significance of using machine learning tools as an approach and technology capable of processing massive and complex data, as well as identifying nonlinear relationships between financial and operational variables. These models offer enhanced prediction capabilities for net income determinants, hence opening the door to larger banks to improve profitability and performance. The research has five chapters. The first chapter introduces the chapter, and the second chapter discusses literature review and past studies into the determinants of net income and statistical models/machine learning techniques. Chapter 3 provides the study methodology and chapter 3 provides results of analysis and comparison between traditional models and machine learning algorithms. The study concludes with the chapter five and carries the conclusion, recommendations and suggestions for future research.

The ongoing and dynamic transformation of the banking and financial services landscape has made the utilization of AI technologies a core driver driving both operational efficiencies as well as remaking the very fundamentals of existing practices, and altering the very paradigms of competition and value in the financial services sector ([22]). With the ongoing digital revolution and rapid transformation of the banking and financial services world, the deployment of AI technologies has been an influential element of change as well, not only to enhance operational efficiency, but also to alter the foundation of standard businesses to a new basis and change the competitive and value relationships of the financial world ([23]). Not only do financial implications impact strategic investment decisions, resource allocation efficiency, etc., but measuring the financial sector’s adaptability to digital developments is also important. In the age of unprecedented tech changes and digital transformation that leads to reshaping of financial markets, knowledge of the influence of AI on financial performance, risk, and profitability becomes an essential strategic concern for both industry (financial) investors, government policymakers, and risk managers ([28]; [29]). A wealth of literature about factors that influence banks financial performance has emerged through the course of the past 40 years in the contemporary financial literature. Various internal and external drivers have been recognized as critical to determining factors of profitability and financial stability. Most significantly, asset quality and amount of non-performing loans (NPLs) are regarded as important indicators of the effectiveness of the credit risk management and are among the most important factors that have either a positive or negative impact on banks’ net income or on their operational efficiency indicators ([24]). Recent research has verified the importance of AI-based virtual assistants to increase banks’ operational profitability, enhance the efficiency of customer service, decrease the need for classical human resources through their abilities to automate repetitive processes, and generate timely and correct solutions. This change indicates a transformation in how institutional performance is being performed in the banking sector, with AI becoming not more of a supporting tool but a strategic player which is redefining business models and financial performance measurements ([7]). In a study covering 23 European banks over the period (2010-2019), using a fixed effects panel regression model adjusted for unobserved fixed effects, a positive and statistically significant association was found between the adoption of advanced digital technologies and improved financial indicators, including return on assets (ROA) and return on equity (ROE). The evidence suggests that the gradual acceptance of technological innovations is a crucial element that leads to higher efficiency and better financial performance in the banks in Europe ([1]). Return on assets (ROA) has been employed as a recognized and widely used metric within the financial literature to measure the financial performance of banks, as it is able to assess the efficiency of resource management in turn producing a profit without the dependence on the financing structure, thus proving an appropriate predictor for benchmarking different financial institutions with varied capital assets and leverage ratios ([2]). The significant development and real-life practice of AI started to gather steam in the 21st century due to powerful cybercomputing and huge datasets ([19]). The pace of new AI technologies seems to accelerate, potentially changing the terrain of voluntary disclosure in banking; for banks will be able to sift through huge volumes of unstructured and disconnected data, identify and quantify performance indicators that are simply wrong, or perhaps not that good, thereby creating a more accurate and open-source form of information that is beyond just reporting a claim—opening up entirely new standards of transparency and accountability in financial reporting ([21]). The rivalry in banking sector is a significant consideration when it comes to voluntary decision-making for firms to provide of AI-facilitated technologies—organizations generally are encouraged to increase the transparency of strategic data as a strategic advantage on offer, to boost investor and customer confidence, on the one hand, and to obtain competitive advantages on the other hand. At this juncture, banks wanting to differentiate themselves in an oversaturated market have clearly shown a bias in disclosing their AI success and initiatives—to perform not just regulatory requirements, but also to differentiate themselves in their institutions as well as demonstrate technical competence and progress ([27]). AI technology use in financial services and manufacturing is gaining significant traction and the data-based approaches behind predictive analytics have come to dominate the business field with respect to advanced AI. Banks that align their ethical values by practicing transparency, making corporate social responsibility, and ethical behaviour are among the clients most likely to disclose the usage of AI systems voluntarily ([6]). Especially when banks see the disclosure as a tool to increase institutional trustworthiness and credibility among key stakeholders and it is a reputation management strategy, also for the financial institutions being ethical sustainability.

Its methodology involves an empirical approach, quantitative survey of financial data and advanced machine learning application aiming at finding determinants of banks’ Net Income and predicting the same better than traditional models ([26]). The methodology of this study is divided into five major steps: definition of study variables and collecting data. Based on previous theoretical literature, the selected financial and operational indicators that are among the crucial determinants of net income were:

Interest Income (II);Interest Expense (IE);Average Earning (AE);Net Income (NI);Total Assets (TA);Shareholder Equity (SE);Operating Expenses (OE);Operating Income (OI);Market Share (MS). The quantitative data were collected at the level of major commercial banks in the United States of America over the period of 3 January 2022 to 23 March 2023. These quantitative data were obtained from the Federal Deposit Insurance Corporation. Outliers and missing values were removed from the data. In this period witnessed exceptional economic conditions, most notably the US Federal Reserve’s sharp interest rate hikes to combat high inflation, which led to significant fluctuations in interest rate margins, financing costs, and borrower behavior. Therefore, analyzing data during this period provides a robust test of the models’ predictive capabilities in volatile environments ([10]).

Data Preparation & Processing. The raw data were examined and pre-processed and fixed effects and random effects were also analyzed using panel data. Below we performed the normalisation on this data using the Min-Max Scaler. Derived variables are volatility indexes and annual percentage change (YoY Growth) ([3]). The samples were split down into two data sets: training set 70% and test and validation set 30%. Time-Series Cross-Validation was applied to prevent data leakage to the training data.

Machine Learning Model Development

In this study, we applied some machine learning models in comparison to a conventional economic model ([4]). Model 1: Decision Tree (DT)-Model, which is utilized for the estimation of various variables and does nonlinear prediction without strict distributional assumptions ([5]). Multiple linear regression with fixed effects (GLM), as a traditional standard model used for comparison ([12]).

Model performance evaluation. Several performance indicators were calculated to assess models’ accuracy: MAE (Mean Absolute Error), RMSE (Root Mean Squared Error), and Adjusted R². MAPE (Mean Absolute Percentage Error). ([25]) This approach allows scientific integration between the theoretical world of finance and the technological world of artificial intelligence, and also ensures a replicable framework for further studies on intelligent financial analysis in a practical banking context.

Figure 1 presents a descriptive analysis of the distribution of a set of key financial indicators in the banking sector using boxplots. These plots show the main statistical characteristics of each variable, including the median, the interquartile range (IQR), which covers values between the first (25%) and third (75%) quartiles, and outliers that fall outside the standard limits (i.e., beyond 1.5 × IQR). The variables represented include net income, interest income, operating expenses, total assets, shareholders’ equity, operating income, market share, and average earning assets. This graphical presentation aims to reveal the degree of dispersion, the presence of outliers, and the direction of the distribution (symmetry or skewness), which is a crucial preliminary step for understanding the behavior of the data before applying forecasting models such as the generalized linear model (GLM) or decision tree.

(a) (b)

(c) (d)

(e) (f)

(g) (h)

(i)

Figure 1. Descriptive analysis of the distribution of financial variables using box plots.

The results of the statistical analysis of studies on the financial variables in Figure 1 show clear discrepancies in the performance patterns, which align well with the financial literature on the dynamics of banking and economic indicators. According to the data interest income ranges between 2,000,000 and 3,500,000 with substantial variations, while interest expenses showed high volatility and frequent spikes. This means fluctuations in the cost of financing are due to varying pressures, consistent with the finding ([20]), that the sensitivity of banks to the level of interest rates. On the other hand, average profits were relatively stable, with mid-range centered to a narrow degree with negligible extremes. It is indicative of the stability of operational performance such as ([16]). In keeping with the observation that profit stability is indicative of operating effectiveness. Net income between $0 and $4,000,000 with certain variations fluctuated on a periodic basis, suggesting that profitability in the commercial and financial industry had a cyclical aspect of the financial industry as cited in ([8]). Collectively and in relation, the results showed that total assets are concentrated with a value between $1 million and $3 million, which signifies that a stable asset base is being constructed. On the other hand, the shareholders’ equity looked more stable and less volatile in comparison to the other indicators in ([9]), reflecting the operating activities, both operating expense and profit were relatively stable at between $1 million and $3 million, showing a close correlation between the respective revenues and the corresponding costs. This corroborates ([15]) claim that the income/operating expenses ratio is a key indicator of how effectively management is aligned. Lastly, differences in market share highlighted vast variations in market size due to high rivalry and continuous shift of market structure confirms what [18] stated about the dynamics of competitive markets.

Table 1. The basic descriptive measures.

Table 1 presents the preliminary descriptive measures as the basic descriptive features of the data for nine observations for the period 2022 to 2023 of N = 521 for a sample comprising nine individuals and a simple descriptive model was provided to assess the relations between the IVs and NI. These measures display central distribution features and the dispersion of the variables and give confidence intervals. Almost all the variables follow normal distribution around the mean to show some type of heterogeneity in performance by unit (firms or countries). Outliers are few, as the extremes are more or less consistent with the ones in the other scale. Standard deviations for both shareholders equity (SE) and interest income (IIR) as a measure of profit show reasonably low standard deviations when compared to the mean, suggesting the importance of such a stability for these indicators. Net income (NI) and operating income (OI), however, have a higher variation, showing different levels of profitability according to units. Total assets (TA): between 80 million and 100 million, the average of 89.8 million, that indicate there are fairly large entities included in the sample. Market share (MS) ranges from 2% to 29%, mean 19.37%, reflecting higher competition and large market power of the companies. Operating expenses (OE) comprise a somewhat minor fraction of revenue which could imply successful control. Operating income (OI) is comparatively high, a sizable portion of income, reflecting good operational efficiency.

Figure 2. Decision tree digramm.

This diagram 2 describes a decision tree structure to predict the net income from a series of financial variables. As an analytical technique, this tree is used for understanding the correlation between the financial factors on the profitability of an organization and making a choice and divide the decision through the utilization of statistical conditions e.g., adjusted P-value and F-test ratio. The decision tree indicates interest expense has a significant impact on net income with the first split based on a value of 96,888 (P < 0.001). In a low-interest expense group, shareholder equity is the next key determinant, segregating companies with decent capital among others. This progression makes the obvious distinction that financial soundness through the substitution of less leverage and higher equity is the foundation to the sustainable profitability.

Figure 3. The predictions of the decision tree (DT) model and the generalized linear regression model (GLM).

The DT vs GLM for net income (NI) comparison graph (Figure 3) shows a chart comparing DT vs GLM for forecast performance. This shows the GLM model generates variable predictions that look like random movement, suggesting overfitting and being unable to track the trend of the entire model. The DT model can be expected to show more regular behaviour as compared to the GLM model. Thus the above results emphasize the need to adopt pruning techniques or obtain a simple but robust model on data, given that the data is very volatile.

Table 2. The performance of the Decision Tree (DT) and Generalized Linear Regression (GLM).

The comparison Table 2 of performance of the Decision Tree (DT) and Generalized Linear Regression (GLM) models reveals an obvious and considerable superiority of the Decision Tree model with respect to predictive accuracy for all their performance metrics. The DT model achieves 3x better predictive accuracy with the mean absolute percentage error (MAPE) of 6.79% compared to the GLM 20.42%, indicating it is better able to capture the true pattern of net income. This is further supported by the SMAPE measure, where it demonstrates a similar observation with DT achieving 11.93% as compared to 31.15% for the GLM, indicating DT excels with regard to relative change and volatility without bias. The MASE metric also indicates that both models are better than the naive expectation (where the value is less than 1), but the DT model performs super well with 0.0679, outperforming the reference model by more than 93% compared to 79.6% for the GLM. In particular, this outstanding performance of the decision tree results from its ability to model more complex nonlinear relationships and interactions between variables (which linear models such as GLM are unable to effectively represent or fit), particularly during economic shocks. Therefore, the performance of the metrics is in line with the hypothesis that the decision tree model is the best model for accurate prediction of the NET INCOME dynamics. The poor performance of the GLM model reflects the limitations of linear assumptions in capturing dynamic patterns and interactions between financial variables, resulting in a decrease in its predictive power.

This study represents a qualitative contribution to the field of corporate finance and economic analysis. It seeks to understand the relationship between financial and investment indicators, with a focus on developing accurate predictive models for net income. This interest stems from the critical importance of net income as a key measure of a company’s profitability, financial stability, and competitiveness, especially in volatile economic environments. The study relied on a wide and diverse dataset, including key financial indicators, to build and compare advanced predictive models. The results showed that the Decision Tree (DT) model significantly outperforms the Generalized Linear Regression Model (GLM) in predictive accuracy across all performance metrics. This superiority indicates that machine learning models are capable of capturing nonlinear patterns and complex interactions between variables such as the impact of operating income at certain levels of market share or operational efficiency that traditional statistical models cannot efficiently detect. The practical importance of the results lies in the fact that the decision tree model identified interest expenses and shareholder equity as the most prominent factors affecting net income, which indicates that improving the efficiency of the financing structure and strengthening the capital base can be among the priorities of financial management to achieve higher profitability stability. Through the graph comparing the predictions to the actual values, it is clear that: The GLM model provides relatively stable predictions, but fails to track large fluctuations. While the DT model shows greater accuracy in capturing peaks and troughs, despite some overfitting, which can be subsequently addressed through pruning or the use of hybrid models such as Random Forest. This study presents a new methodological framework that combines traditional quantitative analysis with modern machine learning tools. Accordingly, the study recommends the use of hybrid models that combine GLM and DT to maximize forecast accuracy while maintaining interpretability.