Attention-Augmented GRU for Stock Forecasting: A Trade Off Between Directional Accuracy and Price Prediction Error

Abstract

Attention mechanisms have been widely incorporated into recurrent neural network architectures for financial time series forecasting, with most prior work reporting improvements in pricelevel error metrics. This study revisits that claim through a controlled empirical comparison of four deep learning architectures on nearly two decades of Telkom Indonesia (TLKM) closing price data from the Indonesia Stock Exchange (IDX). The models evaluated are a three-layer Gated Recurrent Unit (GRU) baseline, a comparable Long Short-Term Memory (LSTM) network, a Bahdanau endattention GRU (Attn-GRU-V2), and a multi-head self-attention GRU hybrid (Attn-GRU-V3). Each architecture is trained over 30 independent runs with distinct random seeds, and performance is reported as 95% confidence intervals derived from the t-distribution. Statistical comparisons employ the Wilcoxon signed-rank test, a nonparametric paired test appropriate given the confirmed non-normality of residuals. The main finding is a consistent trade-off: the plain GRU achieves the lowest RMSE (94.02 ± 1.22 IDR) across all 30 runs, while Attn-GRU-V2 achieves the highest directional accuracy (45.91 ± 0.09%), surpassing GRU in every independent run. Bahdanau attention weights are nearly uniform across the 30-day lookback window (coefficient of variation: 3.21%), indicating that the mechanism cannot identify selectively informative timesteps in this univariate price series. This finding is consistent with the weak-form Efficient Market Hypothesis for the Indonesian market. An ablation study reveals that a 20-day lookback window maximizes directional accuracy (47.72 ± 0.21%) for the Attn-GRU-V2 model. These results suggest that Bahdanau end-attention consistently and significantly improves directional accuracy relative to a plain GRU baseline, providing an architecturally attributable advantage for direction-based applications, even when absolute price-level error is not reduced. The directional accuracy values remaining below 50% across all models are consistent with a weak-form efficiency characterization of the Indonesian market.