Predicting corporate bankruptcy remains a core challenge in financial risk management. Traditional models, heavily reliant on static financial ratios, often fail to capture the nuanced and forward-looking signals embedded in unstructured corporate disclosures. This chapter proposes an enhanced approach that integrates machine learning techniques with textual analysis to improve both the accuracy and interpretability of bankruptcy prediction. We leverage firm-level financial reports—specifically the Management Discussion & Analysis (MD&A) sections—to extract meaningful textual features that complement numerical data. Using topic modelling techniques such as Latent Dirichlet Allocation and Dynamic Embedded Topic Models, we identify latent thematic structures that reflect managerial concerns over liquidity, innovation, regulation, and macroeconomic uncertainty. These topical features evolve in sync with major economic events, providing early warning signals of financial distress. When combined with traditional z-score indicators, these features significantly improve classification performance, especially in predicting rare default events. In parallel, we develop a sentiment-based classifier using a domain-specific financial dictionary to quantify tone at the word level. These sentiment scores are shown to enhance one-year-ahead predictive accuracy and offer more intuitive explanations of firm risk than deep learning approaches. Our experiments demonstrate that textual features—both topical and sentiment-based—provide predictive power comparable to or exceeding that of numerical indicators alone. By integrating textual analysis into machine learning workflows, this research not only improves bankruptcy prediction outcomes but also enriches our understanding of corporate behaviour under stress. The findings highlight the potential of unstructured data in credit modelling and suggest promising directions for future work, including multimodal learning and fairness-aware financial AI.