Predicting Pakistan Stock Exchange with Deep Learning
A BiLSTM-Attention hybrid model trained on 36,000+ samples across 15 of the most liquid KSE-100 stocks, incorporating 53 technical, macro, and market features. Evaluated against seven baselines with an honest verdict on what works and what does not.
Arham Mirkar
DataLayer — Enterprise Data Infrastructure
Abstract
We present a deep learning system for next-day price prediction on the Pakistan Stock Exchange (PSX), targeting 15 of the most liquid KSE-100 constituents. The model uses a Bidirectional LSTM with multi-head self-attention over 60-day lookback windows, consuming 53 features spanning technical indicators, KSE-100 market context, FIPI foreign investor flows, USD/PKR exchange rates, Brent crude oil prices, and SBP monetary policy signals. Walk-forward cross-validation across 2021-2024 yields 62.4% directional accuracy, matching the simple MA-5 baseline (62.8%) while the naive close-to-close baseline achieves only 51.2%. However, on MAPE the naive baseline decisively wins (1.64% vs 21.89%) due to corporate action artifacts in unadjusted data. We document five confirmed stock splits and bonus issues, propose V2 adjustments, and argue that directional accuracy — not MAPE — is the correct evaluation metric for trading systems.
Key Findings
The BiLSTM-Attention model achieves 62.4% directional accuracy across 15 stocks, beating the naive baseline (51.2%) and matching the MA-5 rule-based strategy (62.8%). It generates +131.5% simulated PnLwith Sharpe 3.53 on the 2025 test set. The model's price-level MAPE is inflated to 21.89% by 5 undetected corporate actions (stock splits and bonus issues) in SYS, LUCK, UBL, MARI, and ENGRO. After adjusting for these events, the true MAPE on the 11 clean stocks is ~2.8%, comparable to the naive baseline.
Data Pipeline
Multi-source data aggregation from PSX historical prices, KSE-100 index, FIPI flows, and three macro indicators — all with proper temporal alignment.
Target Universe (15 KSE-100 Stocks)
53 Input Features
Technical Indicators
SMA, EMA, MACD, RSI, Bollinger Bands, ATR, OBV, VWAP, volume ratios, momentum, candlestick patterns
Market Context
KSE-100 returns, volatility, volume ratio, distance from 52-week high, FIPI daily flows, cumulative flows, flow trend
Macro Features
USD/PKR rate and returns, Brent crude oil prices, SBP policy rate, rate change direction, sector-wise FIPI allocations
Relative Features
60-day rolling beta, relative strength vs KSE-100, returns (1d/5d/20d), log return, volatility (5d/20d)
Macro Feature Lagging
All macro features (USD/PKR, Brent crude, SBP policy rate) are lagged by 1 trading day to prevent look-ahead bias. Sector FIPI allocations are lagged by 1 month. This ensures the model only uses information that would have been available at prediction time in a live trading scenario.
Model Architecture
Bidirectional LSTM with multi-head self-attention and positional encoding, trained with mixed-precision on an NVIDIA Quadro P2000.
Input (batch, 60, 53) | v Input Projection: Linear(53 -> 128) + GELU | v Positional Encoding (sinusoidal, 60 positions) | v BiLSTM (2 layers, hidden=128, bidirectional -> 256) | v Multi-Head Self-Attention (4 heads, dim=256) + Residual Connection + LayerNorm | v Feed-Forward Network (256 -> 512 -> 256) + Residual Connection + LayerNorm | v Last Timestep: (batch, 256) | v FC Head: 256 -> 128 -> 64 -> 1 (next-day close)
All trainable. Xavier initialization with gradient clipping at 1.0.
NVIDIA Quadro P2000 GPU. AdamW optimizer with cosine annealing schedule.
2021, 2022, 2023, 2024 validation windows. No future data leakage.
What the Model Learned
Attention heatmaps reveal the model focuses on the most recent 5-10 days, with decaying attention on older data — consistent with financial time series behavior.
Attention weight distribution across the 60-day lookback window. Brighter regions indicate higher attention — the model naturally learned recency bias.
Prediction Results
Per-stock predictions on the 2025 out-of-sample test set. The model tracks price movements closely on clean stocks but fails catastrophically on unadjusted corporate actions.
Baseline Comparison
Seven baselines from trivial (naive close) to ML (XGBoost, RandomForest, LogReg). The honest verdict: the LSTM did not beat naive on MAPE.
| Model | MAE | MAPE | Dir Acc | PnL | Sharpe |
|---|---|---|---|---|---|
| Naive Close | 4.72 | 1.64% | 51.2% | +40.1% | 1.12 |
| MA-5 | 5.22 | 1.80% | 62.8% | +145.6% | 3.71 |
| MA-20 | 12.03 | 4.10% | 56.2% | +93.7% | 2.42 |
| Linear Regression | 180.85 | 53.82% | 51.9% | +41.8% | 1.12 |
| XGBoost | 181.03 | 54.86% | 50.0% | +15.2% | 0.41 |
| LogReg (Direction) | N/A | N/A | 50.5% | N/A | N/A |
| Random Forest | N/A | N/A | 50.9% | N/A | N/A |
| BiLSTM+Attention (Ours) | 42.21 | 21.89% | 62.4% | +131.5% | 3.53 |
Honest Verdict
MAPE: Naive wins (1.64% vs 21.89%) — the LSTM learned "tomorrow is close to today" but was contaminated by corporate actions. Direction: LSTM shows real edge at 62.4%, matching MA-5 at 62.8%. The model learned directional signals even though it predicted wrong price levels.
Corporate Action Audit
Five confirmed corporate actions were distorting model predictions. We built an automated audit and back-adjustment pipeline to fix the data for V2.
| Stock | Event | Date | Impact | Status |
|---|---|---|---|---|
| SYS | 5:1 Stock Split | Mar 2025 | Price dropped from ~500 to ~100 in raw data | Back-adjusted |
| LUCK | 1:1 Bonus Issue | Jul 2024 | MAPE inflated to 91.25% on unadjusted data | Back-adjusted |
| MARI | Right Issue | 2023 | Minor discontinuity in price series | Back-adjusted |
| UBL | Bonus Issue | 2024 | MAPE inflated to 36.54% on unadjusted data | Back-adjusted |
| ENGRO | 1:1 Bonus | 2021 | Historical price halved at split date | Back-adjusted |
Per-Stock Breakdown
Directional accuracy is above 60% on 10 of 14 stocks. MAPE is under 4% on the 11 stocks without corporate action contamination.
Stocks with red borders have MAPE inflated by corporate actions (stock splits, bonus issues) in unadjusted training data.
V2 Architecture (In Progress)
Return-Based Targets
Shift from absolute price prediction to next-day log return prediction. This makes the model scale-invariant and eliminates corporate action artifacts entirely.
Multi-Head Architecture
Four output heads: return regression (Huber loss), direction classification (BCE), volatility prediction (Gaussian NLL), and confidence calibration. Composite loss weighted 35/35/15/15.
Corporate Action Adjustment
Automated back-adjustment pipeline that detects stock splits, bonus issues, and right issues from PSX announcements and adjusts all historical OHLCV data accordingly.
High-Confidence Trading Filter
The confidence head produces a 0-1 score. Only trade when confidence exceeds the 80th percentile — backtested with 0.5% round-trip transaction costs to simulate real PSX trading conditions.
Methodology & Disclaimers
Not Investment Advice
This is a research experiment. Past simulated performance does not guarantee future returns. The model has not been tested with real capital, slippage, or market impact.
Simulated PnL Assumptions
PnL figures assume: equal-weight allocation across stocks, instant execution at close price, no slippage, no market impact, and positions rebalanced daily. Real-world performance would be lower after transaction costs (PSX: ~0.5% round-trip including brokerage, CDC, and taxes).
Data Quality
V1 was trained on unadjusted OHLCV data from PSX. Corporate actions (splits, bonuses) were not accounted for in V1 training, leading to inflated error metrics on affected stocks. V2 addresses this with back-adjusted data.
Data Sources & References
- Pakistan Stock Exchange (PSX) — historical OHLCV data, 2016-2026
- KSE-100 Index — daily index values and trading volumes
- FIPI (Foreign Investors Portfolio Investment) — daily net flows by market type
- State Bank of Pakistan (SBP) — monetary policy rate history, 2015-2026
- Yahoo Finance — USD/PKR exchange rate and Brent crude oil prices
- PSX Announcements Portal — corporate action filings (splits, bonus, rights)
- Hochreiter & Schmidhuber (1997) — Long Short-Term Memory, Neural Computation
- Vaswani et al. (2017) — Attention Is All You Need, NeurIPS
© 2026 Arham Mirkar, DataLayer. All data sourced from publicly available APIs and exchanges. This paper is for research and educational purposes only and does not constitute investment advice. Code available on GitHub.