Deep Learning for Stock Market Prediction: Evidence from the Pakistan Stock Exchange A CNN-BiLSTM Attention Hybrid Framework with Technical-Indicator Fusion

Abstract

Forecasting the direction and level of stock prices in frontier and emerging markets remains a difficult problem because of thin liquidity, structural breaks, and recurring macro-political shocks. This paper investigates deep learning approaches for one-day-ahead forecasting of the Karachi Stock Exchange 100 (KSE-100) Index of the Pakistan Stock Exchange (PSX) over a ten-year daily sample (2015–2024) that spans the 2017 political crisis, the 2020 COVID-19 crash, the 2022–2023 balance-of-payments and inflation crisis, and the 2023–2024 post-IMF bull rally. Nineteen price-derived and technical-indicator features (moving averages, MACD, RSI, Bollinger Bands, stochastic oscillators, ATR, and volume) are engineered from raw OHLCV data and fed into a 30-day sliding window. We benchmark a Naive random-walk model, ARIMA(1,1,1), a vanilla LSTM, a GRU, and a CNN-LSTM against a novel CNN-BiLSTM-Attention hybrid that couples one-dimensional convolutional feature extraction with a bidirectional LSTM temporal encoder and an additive attention pooling layer, enabling both accurate forecasting and interpretable identification of the most informative lags in the input window. All deep models are trained to predict the standardized log-return rather than the raw price level, which substantially improves robustness to the strong regime shifts present in the KSE-100 series. Experimental results on a chronologically held-out test set (2024) show that all deep learning architectures reduce RMSE and MAPE relative to ARIMA(1,1,1) and the naive persistence benchmark and achieve directional accuracy of 54–55%, compared with 46.5% for ARIMA and 50% (by construction) for the naive model. The proposed CNN-BiLSTM-Attention model attains the lowest MAPE among the attention-augmented models (1.22%) and provides an interpretable attention profile showing that the model relies most heavily on price action from roughly three to four weeks prior, consistent with the medium-term momentum patterns reported in the Pakistani market literature. The framework, code pipeline, and evaluation protocol are designed to be directly reproducible on live PSX data and are offered as a template for future deep-learning-based forecasting studies on the KSE-100 and other South Asian frontier markets.