2.1 Data

Our sample includes all publicly traded FinTech companies globally with an issuer-level credit rating from Standard & Poor’s (^{S&P Capital IQ, 2024}). The analysis period covers the years 2021 to 2023. In our study, the target variable (dependent variable) is the credit rating, and the features (independent variables) are commonly employed accounting and market measures in the literature. All data were obtained from Capital IQ and ^{Bloomberg (2024)}. Our original database includes 34 credit-rating observations with an imbalanced number of observations in each class. For example, there is only one observation in the database for the AA-, A+, and BBB+ ratings, while the BB- rating has six observations.

Also, our database does not contain observations across the entire S&P rating scale. To solve this problem, ^{Doumpos et al. (2015)} regrouped the companies in their sample into five classes: (1) AA- to AAA; (2) A-, A and A+; (3) BBB-, BBB, BBB+; (4) BB-, BB, BB+; and (5) D to B+. Regrouping credit ratings into classes is problematic because the distance between credit ratings is unknown. In our study, we used, as a criterion to solve this problem, the description of the S&P rating scale (^{S&P Global, 2024}), which groups ratings into three categories based on the ability of issuers to meet their financial obligations. P 1 presents the S&P rating scale and regrouping based on the issuers’ capacity in Table 1 (see Table 1).

Some FinTech companies in our database have a banking license, while others do not. Companies with a banking license must disclose information about their risk exposure. Various studies on bank credit risk have utilized the CAMELS methodology, which stands for Capital Adequacy, Asset Quality, Management, Earnings, Liquidity, and Sensitivity to Market Risk, as a guide for selecting feature variables. Since applying the CAMELS methodology to all FinTech companies in our database is not feasible, we selected the features by researching common financial ratios and financial variables reported in various CRAs methodologies and the literature.

Some studies use only accounting measures, while others use a combination of accounting and market metrics. For example, ^{Galil et al. (2023)} identified that market capitalization and accounting measures, such as interest coverage and dividends, are influential variables in determining credit ratings. ^{Doumpos et al. (2015)} used the distance-to-default obtained from a structural model and accounting data as explanatory variables in a classification model to predict credit ratings. ^{Hajek and Michalak (2013)} found that firm size and market value ratios are the most critical parameters in the United States rating methodology. In contrast, the rating process for European firms relies heavily on profitability and leverage ratios. ^{Jiang (2022)} found that equity risk measures (i.e., beta, alpha, and idiosyncratic risk) account for an increasing share of the total rating variation. In addition to the features used in these studies and those commonly used in the literature, we included valuation measures such as the spread between the cost of equity (measured through the capital asset model, CAPM) and the return on equity (ROE) as well as measures of quality of earnings such as the M-Score (^{Beneish, 1999}), the accrual ratio, and the cash flow from operations (CFO) to net income ratio.

In our study, we focused on variables that were not strongly correlated with one another. For instance, we used the DuPont identity to represent return on equity (ROE) to minimize redundancy. Table 2 displays the features we used to predict FinTech’s credit ratings (see Table 2).

CRAs use a rating-through-the-cycle methodology to capture long-term solvency (^{Kiff et al., 2013}). To reflect FinTech’s permanent economic attributes, we calculated the average of the independent variables used during the analysis period. This criterion is consistent with the one of ^{Hajek and Michalak (2013)}. Table 3 shows the descriptive statistics of the features (see Table 3).

Our sample’s ratios and financial measures show a monotonic relationship with the credit ratings, as described by ^{Metz and Cantor (2006)}.

2.2 Data Preprocessing

The assumptions of traditional statistical models do not bind machine learning algorithms, but it is assumed that the model’s features contribute equally to the prediction of the target variable. The scale of the features influences their predictive importance, so it is necessary to transform them into a standard scale. To improve the efficiency and performance of the algorithms used, we standardized the data with mean 0 and variance 1.

2.3 Resampling

Because of the nature of the problem, credit rating forecasting models use imbalanced datasets, which poses a challenge for model training. In an imbalanced database, the number of observations between classes differs, leading to models with poor predictive performance. This issue arises because most classification algorithms were designed assuming an equal number of observations between classes (^{Brownlee, 2021}). In the literature, two alternatives exist to solve this problem: improving the algorithm or balancing the database (^{Sundar & Punniyamoorthy, 2019}). Ensemble models are generally used to improve the algorithm, combining multiple learning algorithms with low predictive power to improve their accuracy. Data balancing involves using undersampling (i.e., reducing the samples of the majority class) and oversampling techniques (i.e., creating new samples of the minority class).

^{Dastile et al. (2020)} found that only 18% of the top-rated studies have balanced their databases, and that the most common technique is undersampling the majority class.

To solve the class imbalance problem, ^{Chawla et al. (2002)} proposed the Synthetic Minority Oversampling Technique (SMOTE), which creates new synthetic examples of the minority class by joining the nearest neighbors in the feature space. ^{Dastile et al. (2020)} argue that SMOTE is the recommended methodology for imbalanced databases.

As shown in Table 2, Class 3 is the majority class, and Class 1 is the minority class (see Table 2). Balancing the database is critical to avoiding bias towards the majority class. Due to the limited number of observations in our database, employing an undersampling technique may result in the loss of valuable information and subsequently lead to reduced classifier performance. To balance the database, we oversampled the minority class using the SMOTE family of algorithms for R Studio, developed by Wacharasak Siriseriwan (^{Siriseriwan, 2021}). The SMOTE technique allows the creation of synthetic examples of minority classes that are required to improve the predictive ability of machine learning algorithms (^{Brownlee, 2021}). Our study established an equal distribution with 20 observations in each class.

Following standard practice in the literature, we partitioned the database into three groups: 1) training (60%), 2) validation (20%), and 3) testing (20%), to reduce sampling bias. As our sample size is small, we used the resampling technique known as K-fold cross-validation and thus avoided significantly reducing the training set. This technique allows the model to be evaluated multiple times by creating random data combinations and grouping them into K folds with K-1 training samples and one validation sample. We set the parameter K to 10. Table 4 displays the distribution of observations for each credit rating class after synthetic data has been generated (see Table 4).

2.4 Supervised Machine Learning Algorithms Used

a) Classification and Regression Trees (CART)

CART is a supervised machine learning algorithm based on logical conditions to classify or predict data. The technique creates binary trees composed of an initial root node, decision nodes, and terminal nodes representing a single feature (f) and a cutoff value I for that feature. This generates the most comprehensive separation of the labeled data while minimizing the classification error. This error is the decision tree’s impurity function E(T), with the entropy function and the Gini diversity index being the most used. This recursive binary partitioning process continues by forming smaller and smaller subgroups at the decision nodes until the terminal nodes, where labels are assigned to the input data, are obtained. The key hyperparameter of the algorithm is the maximum depth of the tree.

b) K-Nearest Neighbors (KNN)

KNN is a supervised machine learning algorithm used primarily to solve classification problems. This technique classifies a new observation x₀ by finding the K_i instances with the most votes for the nearest neighbors between it and the N training samples x_i,…x_n. Commonly, Euclidean distance is used to find the K_i instances in the training dataset that are most similar to the new observation. KNN classification is generally represented as:

N_k(x₀,d) are the k closest neighbors of x₀ in terms of d, and M is a marker that takes the value of 1 if true and 0 if false.

c) Support Vector Machine (SVM)

SVM is a supervised machine learning algorithm for classification and regression problems. This technique maximizes the distance (margin) between the separating hyperplane and the training data closest to the hyperplane (i.e., the support vectors). Typically, data cannot be perfectly separated by the hyperplane, so the algorithm can be adapted using a soft-margin classification, which consists of introducing a tuning parameter (C) that allows for some errors in the classification while penalizing them. Alternatively, Kernels or non-linear separating boundaries can be used. The SVM algorithm was initially developed for binary classification (an alternative to binary logistic regression), but its application was extended to multiclass classification.

d) Artificial Neural Networks (ANNs)

ANNs are algorithms for classification and regression in supervised and unsupervised learning. Artificial neural networks (ANNs) consist of interconnected nodes known as neurons. These neurons transmit signals from one node to another using a sum operator and an activation function, similar to how nerve impulses are transmitted between neurons in a biological brain. The most common activation functions are the sigmoid and the rectified linear unit (ReLU), which transform the data non-linearly. Information flows through the model via an input layer that receives the data, hidden layers, where learning occurs during training, and an output layer, where the results are obtained.

2.5 Hyperparameter Tuning

Machine learning algorithms involve two errors that cannot be eliminated: bias (related to underfitting) and variance (related to overfitting). The goal is to minimize the total error (bias error plus variance error) by balancing underfitting and overfitting. In practical terms, there is not a standard method for estimating hyperparameters. Instead, heuristics and automated procedures such as grid search, random search, Bayesian optimization, gradient-based hyperparameter optimization, and evolutionary algorithms are among the most used methods for finding the optimal values of hyperparameters.

We employed Bayesian optimization to fine-tune the hyperparameters of the algorithms utilized in our analysis. Bayesian optimization helps identify hyperparameter values that minimize the loss function. Typically, this approach models the algorithm’s capacity to accurately predict outcomes on unseen data using a Gaussian process sample (^{Snoek et al., 2012}). The algorithm samples the input space iteratively, gaining insights from each evaluation until it achieves convergence. Table 5 displays the hyperparameters utilized for training the classification algorithms (see Table 5).