Deep Learning for Stock Price Prediction: A Comprehensive Study on Indian Markets

Abstract

This paper presents a comprehensive evaluation of deep learning architectures for stock price prediction in Indian equity markets. We analyze the performance of Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRU), Transformer-based models, and hybrid CNN-LSTM architectures across all Nifty 50 components over a 10-year period (2016-2026). Our study incorporates multi-modal inputs including price data, technical indicators, fundamental features, and market microstructure variables. Results demonstrate that attention-based Transformer models outperform traditional recurrent architectures for medium-term (5-20 day) predictions, while LSTM models maintain superiority for short-term (1-3 day) forecasts.

1. Introduction and Motivation

The Indian stock market, with a combined market capitalization exceeding $4.5 trillion across NSE and BSE, represents one of the most dynamic and complex financial ecosystems globally. Traditional quantitative models based on linear regression, ARIMA, and GARCH frameworks have shown limited predictive power in capturing the non-linear dynamics and regime changes characteristic of emerging markets. Deep learning, with its ability to model complex non-linear relationships and extract hierarchical features from high-dimensional data, offers a promising alternative.

Our motivation stems from three key observations: first, Indian markets exhibit unique characteristics including higher retail participation, different trading hour dynamics, and sensitivity to domestic policy events that may not be captured by models trained on developed market data. Second, the rapid digitization of Indian markets has generated rich datasets amenable to deep learning approaches. Third, existing literature on deep learning for Indian stock prediction remains fragmented, with most studies focusing on individual stocks rather than systematic evaluation across a broad universe.

2. Literature Review

Recent advances in financial deep learning have produced several notable results globally. Fischer and Krauss (2018) demonstrated that LSTM networks outperform random forests and logistic regression for S&P 500 stock prediction. Ding et al. (2015) showed that convolutional neural networks can effectively extract features from financial news for market prediction. The Transformer architecture, introduced by Vaswani et al. (2017), has been adapted for financial time series with promising results by Li et al. (2022).

For Indian markets specifically, research has been more limited. Patel et al. (2015) compared SVMs and neural networks for CNX Nifty prediction using technical indicators. Gururaj et al. (2019) applied LSTM for Nifty 50 price prediction using only close prices. Our work extends these studies by systematically evaluating modern deep learning architectures across the entire Nifty 50 universe with multi-modal inputs and rigorous out-of-sample testing methodologies.

3. Methodology

Our dataset comprises daily OHLCV data for all current and historical Nifty 50 constituents from January 2016 to December 2025, sourced from NSE India official databases. We engineer 47 features per stock per day: 5 price/volume features, 22 technical indicators (RSI, MACD, Bollinger Bands, ADX, etc.), 8 fundamental features (quarterly P/E, P/B, ROE, debt-equity), 7 market microstructure features (bid-ask spread, delivery percentage, OI changes), and 5 macro features (VIX, USD/INR, FII flows, crude oil, US 10Y yield).

We implement four architectures: (a) Stacked LSTM with 3 layers (128, 64, 32 units) and dropout regularization; (b) Bidirectional GRU with attention mechanism; (c) Temporal Fusion Transformer (TFT) adapted for financial data with variable selection networks; and (d) Hybrid CNN-LSTM combining 1D convolutions for feature extraction with LSTM for temporal modeling. All models are trained using walk-forward optimization with 3-year training windows and 1-year test windows, preventing any look-ahead bias.

4. Feature Engineering and Data Processing

Feature engineering proved critical for model performance. Raw price data undergoes log-return transformation to achieve approximate stationarity. Technical indicators are normalized using 252-day rolling z-scores to maintain scale consistency across stocks and time periods. Fundamental data is forward-filled to handle quarterly reporting gaps and normalized cross-sectionally (within the same quarter across stocks) to capture relative valuation.

We address the challenge of handling corporate actions (splits, bonuses, dividends) through adjusted price series and develop a novel approach for handling missing data in the market microstructure features using a masked attention mechanism that allows the model to learn from available information without imputing potentially misleading values. Input sequences of 60 trading days are used for daily predictions, chosen through cross-validation across sequence lengths of 20, 40, 60, 80, and 120 days.

5. Results and Analysis

Our results reveal architecture-specific strengths across prediction horizons. For 1-day ahead prediction, the Stacked LSTM achieves the highest directional accuracy of 54.3% across the Nifty 50 universe, compared to 53.8% for GRU, 53.1% for TFT, and 52.9% for CNN-LSTM. While these accuracies may appear modest, they are statistically significant (p < 0.01 using binomial test against 50% baseline) and sufficient for profitable trading after transaction costs.

For 5-day and 20-day horizons, the Temporal Fusion Transformer demonstrates clear superiority, achieving directional accuracies of 56.7% and 58.2% respectively, compared to LSTM's 54.1% and 55.3%. The TFT's variable selection network reveals that macro features (VIX and FII flows) gain importance for longer horizons, while technical indicators dominate short-term predictions. Sector-wise analysis shows highest predictability in banking (59.1% at 20-day) and lowest in pharma (52.8%), consistent with the hypothesis that more liquid, sentiment-driven sectors exhibit stronger patterns.

6. Practical Trading Implications

We construct a long-short portfolio using the top and bottom quintile predictions from the TFT model at 20-day horizon. After transaction costs of 0.1% per round trip, the strategy generates a CAGR of 23.4% with a Sharpe ratio of 1.42 and maximum drawdown of 18.7% over the out-of-sample period (2023-2025). This compares favorably to the Nifty 50's 12.1% CAGR with 0.72 Sharpe and 26.3% maximum drawdown over the same period.

However, we note significant time variation in model performance — the strategy underperforms during rapid regime changes (COVID crash, post-election rallies) where historical patterns provide limited guidance. Ensemble approaches combining multiple architectures and horizons show improved robustness, reducing maximum drawdown to 15.2% while maintaining similar returns. These findings suggest that deep learning models are best deployed as one component of a multi-strategy framework rather than as standalone prediction systems.

7. Conclusions and Future Work

This study provides the most comprehensive evaluation of deep learning architectures for Indian stock market prediction to date. Key findings include: (1) Transformer-based models excel at medium-term predictions while LSTMs maintain an edge for short-term forecasts; (2) multi-modal feature inputs significantly improve all architectures compared to price-only models; (3) market microstructure features provide incremental predictive value unique to Indian markets; and (4) ensemble approaches offer the best risk-adjusted performance for practical trading applications.

Future research directions include incorporating alternative data sources (satellite imagery for commodity stocks, web traffic for tech companies, social media sentiment in Indian languages), exploring reinforcement learning for dynamic portfolio allocation, and developing online learning frameworks that adapt to regime changes more quickly. We also plan to extend this analysis to the broader NSE universe beyond Nifty 50, where information inefficiency is likely higher and deep learning models may demonstrate stronger alpha generation potential.