Do Macroeconomic Conditions and GDELT Financial News Sentiment Improve NIFTY 50 Forecasts? A Point-in-Time Out-of-Sample Evaluation

Accurate forecasting of stock market indices remains highly challenging due to the non linear, noisy nature of financial time series. While recent studies demonstrate the efficacy of Long Short-Term Memory (LSTM) networks, many suffer from data leakage and fail to compare against robust baselines. This study rigorously tests whether integrating macroeconomic variables (India VIX, MIBOR, USD/INR, Consumer Confidence Current Index, Consumer Confidence Future Index, and Unemployment) and unstructured Global Database of Events, Language, and Tone (GDELT) financial sentiment can improve Nifty 50 Index predictions. Using a strictly chronological, point-in-time dataset from November 2018 to June 2026 devoid of look ahead bias, this study evaluates single-layer versus multi-layer LSTMs in predicting daily log returns. An ablation study reveals that while a single-layer LSTM (150 neurons) optimized with dropout regularisation achieves a competitive Symmetric Mean Absolute Percentage Error (SMAPE), Diebold-Mariano testing indicates the forecasts are not statistically significantly superior to a naive random walk (p = 0.8194). Furthermore, an out of sample trading simulation demonstrates that the model fails to overcome a 0.1% transaction cost drag, underperforming a Buy and Hold baseline. These findings underscore the semi-strong efficiency of the Indian equity market and highlight the necessity of rigorous statistical and economic validation in applied deep learning research.