A Comparative Analysis of Deep Learning Models for Stock Price Prediction

Abstract

Indonesian equities exhibit high volatility and non-stationary dynamics, making consistent price forecasting difficult under realistic deployment settings. This study presents a comparative benchmark of Long Short-Term Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) for one-step-ahead (t+1) stock price prediction using Walk-Forward Validation (WFV) to preserve temporal causality and avoid optimistic single-split estimates. Historical data are retrieved from Yahoo Finance and modeled in a multivariate OHLCV setting (Open, High, Low, Close, Volume). After missing-value removal, feature standardization, and Min–Max scaling, the series is converted to supervised samples via a sliding window with a lookback of 30 trading days; evaluation is focused on the Close variable. Model performance is assessed using MAE, RMSE, and R², including inter-fold variability to quantify stability across market regimes. Across five Indonesian tickers (AGRO, ADES, ADMF, AALI, ADHI), LSTM consistently outperforms Bi-LSTM (5/5 tickers) in both MAE and RMSE, indicating that the added bidirectional complexity does not translate into improved out-of-sample forecasting under WFV. The best error performance is achieved by LSTM on AGRO (MAE = 26.99, RMSE = 32.72), while the least-negative goodness-of-fit is observed on LSTM AALI (R² = -0.63), suggesting that both deep models may still underperform naïve baselines in several folds. Overall, the results support LSTM as a more stable and implementation-ready benchmark for Indonesian stock forecasting under time-aware evaluation, while highlighting the need for explicit baseline comparisons and stronger feature/target designs to improve out-of-sample generalization