Assessing Hybrid Financial Distress Prediction Models Using Fusion Methods: Evidence from Finnish Listed Firms

Abstract

This thesis evaluates a hybrid approach for forecasting financial stress among Finnish listed companies using quarterly data. The framework combines (i) numerical financial indicators modeled with logistic regression, Extreme Gradient Boosting (XGBoost), and a multilayer perceptron (ANN), and (ii) Management Discussion and Analysis (MD&A) disclosures encoded using FinBERT sentence embeddings and aggregated with a TextCNN feature extractor. Two fusion strategies are assessed: feature-level concatenation and weighted ensembling of predicted probabilities. Using 50 repeated experiments under two training sampling settings (1:1 and 1:3), numerical models provide strong performance at the 3-quarter horizon, with logistic regression yielding stable ranking and separation (AUC/KS). In text-only comparisons, a FinBERT baseline is competitive at 3Q, while the contribution of richer text features becomes more pronounced at the 5-quarter horizon, where FinBERT-TextCNN performs more favorably. Across settings, the preferred fusion strategy depends on both the classifier and the evaluation criterion, reflecting trade-offs between ranking performance (AUC), maximum separability (KS), and fixed-threshold accuracy (ACC). Consistent with prior evidence that direct concatenation can introduce redundancy, weighted ensemble fusion more often yields the strongest overall results, particularly at the longer horizon. Overall, the results indicate that quarterly distress forecasting benefits from combining modalities, with numerical indicators dominating short-horizon predictability and textual disclosures providing complementary value as the forecast horizon lengthens.