NEWS COVERAGE DENSITY AS A BINDING CONSTRAINT ON SENTIMENT-AUGMENTED DEEP LEARNING FOR INDIVIDUAL-STOCK PREDICTION IN EMERGING MARKETS

Abstract

Sentiment-augmented forecasting is celebrated on results drawn mostly from densely covered indices in developed markets; the conditions under which it fails are documented far less often. We make one such condition visible. News coverage density operates as a binding constraint on individual-equity predictive performance, and the limitation surfaces the moment a sentiment model must price a single emerging-market name rather than an index. Habib Bank Limited (HBL), the largest commercial bank in Pakistan and a heavyweight in the KSE-100, should on intuition be among the easiest individual PSX names for a sentiment model to address. A reproducible pipeline assembled 2,503 trading days (January 2016–August 2025), pairing daily prices with 385 date-resolved Dawn News articles (8.4% single-source density), 36 quarterly reports scored by FinBERT-tone aspect-based sentiment, and 10 annual reports across the same aspect families. FinBERT (ProsusAI) produced signed daily scores; a Weighted Aspect Sentiment Index (WASI) summarised each quarterly report across ten financial aspects. Four configurations carried the load: a BiLSTM(64,32) regressor, a BiLSTM(128,64) five-day-horizon model with LayerNorm and multi-head attention, a regularised BiLSTM(32,16) classifier, and a 500-tree Random Forest, with a Logistic Regression ablation control. A multi-source extension folded Business Recorder and Profit by Pakistan Today into the daily channel, raising combined density to 10.51% (263 days). McNemar's test, permutation importance, SHAP attribution, DeLong confidence intervals, calibration analysis, and a long/flat backtest with explicit exposure reporting completed the evaluation suite. Five converging tests document the constraint. The BiLSTM(64,32) regressor returns a test R² of −0.0413 and a Pearson correlation r = 0.017 (p = 0.7607), with directional accuracy of 51.6%. The regularised classifier reaches 50.3% accuracy with AUC 0.5423 (McNemar p = 0.6851); the Random Forest reaches 52.4% (p = 0.4123). Every AUC confidence interval spans 0.5. All five McNemar tests fail to reject the null. Combined three-source coverage of 10.51% remains short of the ≈ 11% SentARL threshold. A long/flat backtest produces a partial-exposure Sharpe artifact: a Sharpe of 2.077 at 29.7% exposure against buy-and-hold's 1.885 at full exposure, even as the strategy's 86.2% cumulative return trails buy-and-hold's 156.4%; the deflated Sharpe ratio removes the apparent edge. In this study the constraint is a Dawn News-specific channel limitation rather than an intrinsic, invariant property of the PSX. Adding two further outlets lifts coverage without crossing the threshold. The implication for practitioners is direct: measure coverage density first, then decide whether sentiment modelling is worth the engineering effort.

Keywords : news coverage density, sentiment analysis, FinBERT, BiLSTM, Pakistan Stock Exchange, negative results, stock-return prediction, partial-exposure Sharpe artifact, SHAP, deflated Sharpe ratio.