Explainable and Responsible AI for Deep Learning-Based NIFTY 50 Stock Price Prediction Using SHAP and Technical Indicators

This paper started with a problem that is more practical than theoretical: deep learning models for stock market prediction work reasonably well, but traders cannot verify why they make the calls they make. We set out to change that for NIFTY 50 prediction, and we think the results make a genuine contribution to that goal. We trained five deep learning architectures on five years of NIFTY 50 data using 41 technical indicators. GRU outperformed the others on every important metric — RMSE of 544.69 INR, Directional Accuracy of 56.82%, and a Composite Score of 0.623. The Diebold-Mariano test confirmed this advantage is statistically robust, not a lucky outcome. LSTM and Bi-LSTM both collapsed to predicting near the mean with zero directional skill, and Transformer and CNN-LSTM both landed below the 50% random baseline on direction. The SHAP analysis on the GRU model yielded findings that are genuinely useful for the field. The 50-day Simple Moving Average is the most important single indicator (Mean |SHAP| = 0.000835), and removing it from the model more than doubles the RMSE. Trend indicators as a group account for 33.61% of the model's decision-making, closely followed by Lag features at 29.24% and Volatility indicators at 28.78%. Momentum oscillators contributed just 4.83%, and volume-based indicators contributed only 1.17% — a finding that challenges the conventional emphasis on volume confirmation in technical analysis. Most importantly, the individual BUY and SELL day waterfall charts demonstrate that SHAP can translate a neural network's internal reasoning into a format that any technically-oriented trader can read and evaluate. On the BUY day (t=15), every major indicator aligned in the same direction. On the SELL day (t=16) that followed immediately, the same indicators lost enough of their conviction collectively to flip the directional call. Neither of those dynamics would have been visible without SHAP. They are now. For future research, we suggest three directions. First, applying the same SHAP framework to individual NIFTY 50 constituent stocks to see whether the indicator rankings change at the stock level. Second, incorporating temporal SHAP — keeping the time axis instead of averaging over it — to identify which of the 60 days in the look-back window contribute most to the prediction. Third, testing this approach in bearish and sideways market periods to see how the category attribution shifts when trend-following conditions break down. The code and configuration for this study are available from the corresponding author on request. We hope the methodology described here — clear data pipeline, reproducible evaluation, and SHAP-based transparency at global, category, and day-specific levels — is useful to researchers working on explainable AI in other financial contexts.

[1]  S. Hochreiter and J. Schmidhuber, "Long short-term memory," Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997. [2]   K. Cho et al., "Learning phrase representations using RNN encoder-decoder for statistical machine translation," in Proc. EMNLP, 2014, pp. 1724–1734. [3]  M. Schuster and K. K. Paliwal, "Bidirectional recurrent neural networks," IEEE Trans. Signal Process., vol. 45, no. 11, pp. 2673–2681, 1997. [4]  A. Vaswani et al., "Attention is all you need," in Advances in Neural Information Processing Systems, vol. 30, 2017, pp. 5998–6008. [5]   S. M. Lundberg and S.-I. Lee, "A unified approach to interpreting model predictions," in Advances in Neural Information Processing Systems, vol. 30, 2017, pp. 4765–4774. [6]  T. Fischer and C. Krauss, "Deep learning with long short-term memory networks for financial market predictions," European Journal of Operational Research, vol. 270, no. 2, pp. 654–669, 2018. [7]   F. A. Gers, J. Schmidhuber, and F. Cummins, "Learning to forget: Continual prediction with LSTM," Neural Computation, vol. 12, no. 10, pp. 2451–2471, 2000. [8]  I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. Cambridge, MA: MIT Press, 2016. [9]   F. X. Diebold and R. S. Mariano, "Comparing predictive accuracy," Journal of Business and Economic Statistics, vol. 13, no. 3, pp. 253–263, 1995. [10]  J. J. Murphy, Technical Analysis of the Financial Markets. New York: New York Institute of Finance, 1999.