Comparative Evaluation of Deep Learning Architectures for Next-Day Stock Price Forecasting Using Technical Indicators

Abstract

Accurate next-day stock price forecasting remains challenging because daily price changes have a low signal-to-noise ratio and can be strongly affected by short-lived news shocks, order-flow imbalances, and abrupt changes in volatility or market sentiment. This study presents a controlled empirical comparison of deep learning architectures for next-day stock price forecasting using technical indicators. Using a decade-long daily dataset covering four large-cap NASDAQ equities (AAPL, META, SBUX, and TSLA), multivariate input sequences are constructed by combining historical prices with five widely used technical indicators: exponential moving average (EMA), relative strength index (RSI), moving average convergence divergence (MACD), on-balance volume (OBV), and average true range (ATR). Four deep sequence architectures—long short-term memory (LSTM), bidirectional LSTM (BiLSTM), gated recurrent unit (GRU), and convolutional LSTM (ConvLSTM)—are evaluated across multiple lookback windows (5, 15, and 30 trading days) and chronological train/validation/test splits (60–20–20, 70–15–15, and 80–10–10). Hyperparameters are optimized through random search, and forecasting performance is assessed on held-out test sets using normalized-scale root mean squared error (RMSE) and out-of-sample R2. Within the examined fixed chronological partitions, ConvLSTM records the lowest observed RMSE for all four equities, attaining values between 0.0256 and 0.0394 and out-of-sample R2 values above 0.90. Because the evaluation does not include walk-forward validation or formal statistical significance testing, these results should be interpreted as descriptive evidence within the present experimental setting rather than as proof of general architectural superiority. To assess practical utility, forecasts are translated into a transparent long-only trading rule that enters the market when the predicted next-day closing price exceeds the current closing price. Out-of-sample backtesting shows that the frictionless forecast-driven strategy achieves higher terminal cumulative returns than Buy-and-Hold for AAPL, SBUX, and TSLA, while Buy-and-Hold remains superior for META. Approximate five-day-frequency risk-adjusted estimates generally reinforce these relative patterns: the ConvLSTM strategy improves the Sharpe, Sortino, and Calmar ratios for AAPL, SBUX, and TSLA, although TSLA remains exposed to substantial drawdown risk. Transaction-cost sensitivity analysis further indicates that the terminal-return gains weaken under trading frictions and are particularly sensitive for AAPL. The findings demonstrate the value of evaluating forecasting architectures through both statistical and financial criteria, while emphasizing that lower point-forecast error does not necessarily translate into superior economic or risk-adjusted performance.