Integrating Disclosure Tone and Machine Learning for Financial Performance Prediction

Abstract

This study aims to enhance financial performance forecasting by integrating disclosure tone variables—positivity, negativity, and uncertainty—into machine learning models for non-financial firms in Pakistan. Existing financial forecasting models often rely solely on quantitative metrics, neglecting the qualitative aspects of corporate disclosures. This study addresses this gap by incorporating tone variables alongside key financial indicators such as ROA, ROE, and EPS. Textual data were extracted from annual reports and analyzed using a financial tone lexicon. After pre-processing, tone scores were combined with financial metrics to train predictive models using Support Vector Machine (SVM), Artificial Neural Networks (ANN), Random Forest (RF), and Long Short-Term Memory (LSTM). Random Forest consistently outperformed other models, achieving the highest accuracy (EPS: 0.82, ROA: 0.82, ROE: 0.85), while SVM and ANN showed moderate performance, and LSTM performed comparatively lower (EPS: 0.64). This research introduces a novel hybrid forecasting approach that integrates qualitative tone analysis with quantitative data, addressing limitations in traditional models. It offers practical insights for investors, analysts, and policymakers operating in emerging markets marked by information asymmetry, and contributes to advancements in financial technology and corporate disclosure practices.