Explainability vs. Accuracy in Financial Distress Prediction: A Multi-Model XAI Analysis of Indian Listed Firms (2011–2025)

Abstract

Predicting financial distress is a longstanding challenge in corporate finance, credit risk management, and regulation. Machine learning models are often extremely accurate, but they are hard to interpret and the most accurate models are often least transparent. This trade-off is particularly troublesome when there is a lot at stake, such as credit evaluation. This thesis is based on an empirical study of the same trade-off by utilising a large panel CMIE Prowess dataset of Indian listed companies from 2011 to 2025. Four models are trained, including Logistic Regression (LR), Random Forest (RF), XGBoost (XGB), and a Multilayer Perceptron (MLP) on ten financial ratios that were selected on theory. The main test uses a temporal split (2011–2019 training, 2020–2025 testing). Financial distress is characterized by early-warning accounting indicator (interest coverage ratio less than one) - this encompasses not only formal default, but financial deterioration. Performance measurement is done using AUC-ROC. There are four model classes with 9 different methods of explanation: coefficient weights, SHAP, permutation importance, LIME, PDP, ICE plots, LOFO. The results show that nonlinear models (RF, XGB, MLP) achieve substantially higher predictive accuracy than Logistic Regression (ΔAUC ≈ 0.10). However, feature attributions diverge systematically between linear and nonlinear model classes, while explanation methods show high consistency within nonlinear models. Trade-off between accuracy and interpretability is not perfect

reliability of explanation will depend on what model and what explanation method is applied. In the case of XGBoost, SHAP and LOFO rankings have moderate, but uncertain agreement (ρ = 0.68, 95% CI includes 0) i.e. black-box models can be explained with conditional reliability when validated using faithfulness checks, but further evidence is needed. But the apparent importance of a feature can be inverted by the choice of the method of explanation, rather than by the model. Interpretability is not solely an attribute of the model, but rather of the model-method combination. Clustering SHAP values on the full test set (n = 32,778) reveals three stable firm types: robustly healthy (~0.4% distress), a vulnerable middle group (~58% distress), and near‑certain distress (~99% distress). These findings may provide evidence for Indian regulation. SEBI and RBI require increasing numbers of model explanations that can be audited. The findings showcase that a singular explanation method is insufficient for compliance. Instead, an XAI audit pipeline should have checks for faithfulness (e.g., the SHAP-LOFO correlation) and tests for temporal stability. Keywords: financial distress prediction, explainable artificial intelligence, SHAP, XGBoost, Indian capital markets, SEBI, model interpretability, temporal validation