Do Technical Indicators Improve Deep Learning Forecasts? An Empirical Ablation Study Across Asset Classes

Technical indicators derived from historical price data have long been central to quantitative trading strategies, yet their actual contribution to modern deep learning forecasting models remains an open empirical question. This study presents a large-scale ablation analysis examining whether technical indicators improve next-day price prediction when used as inputs to recurrent neural networks. We conduct 500 controlled experiments across 10 assets spanning five asset classes—commodities (Crude Oil, Gold), cryptocurrencies (Bitcoin, Ethereum), equities (Apple, Microsoft), foreign exchange (EUR/USD, USD/JPY), and market indices (NASDAQ, S&P 500)—using daily OHLCV data from 2010 to 2025. Five feature configurations are evaluated: a raw OHLCV baseline and four indicator-augmented variants incorporating momentum (RSI, Stochastic Oscillator), trend (SMA, EMA, ADX, MACD), volatility (ATR, Bollinger Bands), and a combined all-indicator set. Each configuration is tested with both LSTM and GRU architectures across five random seeds to ensure statistical robustness. Our results show that technical indicators do not improve—and frequently degrade— forecasting performance relative to raw price data. The baseline OHLCV configuration achieves the lowest mean RMSE (0.166 ± 0.148) and highest mean directional accuracy (55.7% ± 5.5%). Every indicator-augmented configuration produces higher prediction error, with the comprehensive all-indicators variant exhibiting statistically significant degradation (34.6% RMSE increase, p < 0.001, Cohen’s d = −0.29). All four indicator categories show significant performance reduction at α = 0.05. GRU models achieve marginally higher directional accuracy than LSTM (55.3% vs. 51.0%), although RMSE differences between the two architectures are not statistically significant (p = 0.846). Foreign exchange stands out as the only asset class where volatility indicators improve performance (4.2% RMSE reduction), while high-volatility assets (cryptocurrencies, commodities) exhibit 83% higher mean prediction error than their low-volatility counterparts. These findings suggest that deep recurrent architectures implicitly learn the patterns captured by conventional technical indicators, making explicit indicator features redun- dant or even harmful. The results carry practical implications for feature engineering in neural network-based trading systems and highlight the importance of rigorous baseline comparisons in applied financial machine learning.