This study aims to examine whether investor emotions embedded in financial news can predict next-day (t+1) stock market movements in Pakistan. Specifically, it assesses whether incorporating emotion-based indicators, operationalized using the NRC Emotion Lexicon (v0.92), which maps tokenized words to eight primary emotions, into deep learning models improves the accuracy of stock price forecasts for the KSE-100 index on the Pakistan Stock Exchange (PSX).
Financial news articles published between January 2013 and August 2025 were collected from five leading Pakistani English-language news outlets. Investor emotions were quantified by applying the NRC Emotion Lexicon to preprocessed article text, producing normalized emotion scores for eight categories (anger, anticipation, disgust, fear, joy, sadness, surprise and trust) at the article level. These scores were aggregated into daily emotion indices and integrated with historical stock market data as input features for four deep learning architectures: long short-term memory (LSTM) networks, deep neural networks (DNN), convolutional neural networks (CNN) and gated recurrent units (GRU). Model performance was evaluated across multiple forecast horizons (t+1, t+20, and t+50 days) using a rolling-window validation approach, with mean absolute error (MAE) and mean absolute percentage error (MAPE) as the primary evaluation metrics. Granger causality tests were additionally employed to investigate the causal ordering between emotion indices and stock returns.
The results reveal a dominance of negative emotions in financial news, with fear and sadness exerting a significant influence on adverse market movements. Emotion-augmented models outperform baseline models that rely solely on historical price data, with LSTM achieving the highest predictive accuracy (MAPE = 1.58%) at the t+1 horizon. Rolling window validation confirms model robustness across different market regimes. Granger causality tests reveal that disgust, sadness and joy significantly Granger-cause stock returns, while returns in turn Granger-cause anticipation and disgust, indicating bidirectional feedback loops. These findings constitute evidence against the semi-strong form of the efficient market hypothesis (EMH) in the Pakistani market.
This study contributes novel evidence from Pakistan's emerging market by integrating emotion-level (rather than binary positive/negative) sentiment measures with deep learning models for stock prediction. It provides the first systematic test of whether emotion signals in financial news predict stock returns on the PSX, offering evidence on the semi-strong-form efficiency of the Pakistani equity market. The multi-horizon and rolling window evaluation framework offers methodological improvements over prior single-split approaches.
