This study investigates the predictive power of textual data from MD&A reports and conference call transcripts, combined with machine learning algorithms, for predicting discretionary accruals. Advanced text mining techniques, including TF-IDF, LDA, and sentiment analysis, extract meaningful features from textual data to complement traditional financial metrics. Various machine learning models, such as SVR, KRR, RF, and MLP, are evaluated for their predictive performance. The results indicate that textual data, especially from earnings calls, often outperforms MD&A reports. Sentiment analysis provides limited additional predictive power. SVR and KRR consistently perform better, while Decision Trees struggle with complex, industry-specific patterns. This study emphasizes integrating textual data with machine learning techniques to improve financial prediction models. Future research could explore alternative textual sources and broader datasets to enhance predictive accuracy further.