Deep Learning for Emerging Market Forecasting: LSTM Performance on Indian IT Sector Stocks with ARIMA Benchmark Comparison

Abstract

 Forecasting stock prices in emerging markets presents unique challenges rooted in structural volatility, episodic macro shocks, and non-stationary return distributions. While classical econometric approaches such as ARIMA have served as the standard baseline for decades, the question of whether deep learning architectures offer a measurable and reliable advantage in such environments has not been conclusively answered—particularly for individual stock-level prediction in India’s information technology sector. This paper investigates whether a stacked Long Short-Term Memory (LSTM) deep learning model can outperform an ARIMA benchmark when applied to ten years of daily closing price data for five NSE-listed Indian IT stocks: Infosys (INFY), HCL Technologies (HCLTECH), Tech Mahindra (TECHM), Wipro (WIPRO), and Tata Consultancy Services (TCS), covering January 2016 to June 2026. A two-layer LSTM with dropout regularization was trained on 80/10/10 chronological train-validation-test splits, evaluated across three feature configurations—univariate, price-based, and full technical indicator sets—against pre-computed ARIMA forecasts as a benchmark. Seven evaluation criteria were applied: Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Coefficient of Determination (R²), directional accuracy, and Theil’s U statistic. On all five stocks, the LSTM architecture outperformed ARIMA by a wide margin—RMSE fell by 74.1% for INFY, 84.66% for HCLTECH, 87.22% for TECHM, 86.15% for WIPRO, and 88.35% for TCS. The R² values of the best LSTM models spanned 0.77 to 0.94, whereas ARIMA returned negative R² on four of the five series—a clear sign it cannot keep pace with shifting price dynamics. Structured autocorrelation and non-Gaussian residuals nevertheless persisted in the LSTM outputs, pointing to a need for explicit volatility modeling in future work. The results affirm that deep learning provides a strong empirical advantage over classical time series models in emerging market stock forecasting, while also motivating hybrid approaches capable of capturing both trend dynamics and volatility clustering.