Distress Watch: Hybrid Models for Predicting Financial Distress in Companies Listed on the Colombo Stock Exchange

Abstract

Accurately predicting financial distress is essential for managing risk and ensuring stability in financial markets, particularly those of developing economies, where external factors could greatly influence corporate financial health. Traditional models focusing primarily on financial ratios often fail to capture the broader economic context impacting companies' viability. This gap is even more pronounced in the Sri Lankan market, where economic fluctuations strongly affect corporate financial stability. The project attempts to solve the problem by developing a hybrid model that merges company-specific financial data with macroeconomic indicators to enhance the ability to predict distress for companies listed on the Colombo Stock Exchange. A hybrid LSTM-XGBoost model is Proposed to leverage the strengths of both deep learning and gradient-boosting techniques. The LSTM component catches temporal patterns in sequential macroeconomic and financial data, while the XGBoost component boosts capability through feature interaction and nonlinear relationships toward prediction accuracy. Feature scaling and outlier detection were some preprocessing techniques applied in preparing the data. Accordingly, the model was first trained and tested using historical financial and macroeconomic datasets. It employs supervised learning methods to optimize predictive performance. The LSTM-XGBoost hybrid model achieved 87.69% accuracy in predicting financial distress, with strong precision and recall for the ""Alive"" class above 0.88. However, it struggled with the minority ""Failed"" class, resulting in a low F1 score of 0.30. Implementing the Synthetic Minority Over-sampling Technique (SMOTE) improved performance slightly. Future efforts will focus on cost-sensitive learning and improved resampling methods to enhance minority class performance without compromising overall accuracy.