Evaluasi Leakage-Aware dan Imbalance-Sensitive pada BiLSTM dan Machine Learning Klasik untuk Klasifikasi Arah Pergerakan Harga Emas ANTAM

Abstract

ANTAM gold is a widely used hedging instrument among Indonesian investors, yet determining the right moment to transact remains difficult because of its volatile and non-linear price movements. Several prior studies have reported near-perfect predictive accuracy; however, such results frequently stem from evaluation procedures that are prone to data leakage and therefore do not reflect genuine generalization ability. This study develops a leakage-aware and imbalance-sensitive evaluation framework for classifying the directional movement of ANTAM gold prices. Daily price data from 2010 to 2025 (5,751 samples) are transformed into 14 technical features—comprising lagged log-returns, volatility, momentum, moving-average ratios, and RSI—and labelled according to the sign of the five-day forward return. A Bidirectional Long Short-Term Memory (BiLSTM) model is benchmarked against Random Forest, Decision Tree, and a majority-class baseline using five-fold walk-forward validation with purging and train-only feature scaling. Performance is assessed through Balanced Accuracy, Macro-F1, the Matthews Correlation Coefficient (MCC), ROC-AUC, and PR-AUC. All classifiers outperform the majority baseline, with Decision Tree attaining the highest Macro-F1 of 0.534, followed by Random Forest (0.510) and BiLSTM (0.497), and a best MCC of 0.074. These findings indicate limited but real directional predictability and confirm that rigorous evaluation yields markedly more conservative and credible performance estimates than the inflated accuracies claimed in earlier work.