<![CDATA[This study investigates regime-sensitive Value-at-Risk (VaR) forecasting for ANTAM Indonesian gold using a Quantile LSTM under a rigorous walk-forward evaluation protocol. Daily prices are transformed into log-returns, and volatility- and momentum-based features are constructed in a time-respecting manner to prevent information leakage. The model is trained with multi-quantile pinball loss to estimate one-day-ahead VaR at and . Out-of-sample performance is evaluated using a walk-forward expanding-window scheme with periodic retraining, and stress-regime diagnostics are conducted via rolling-volatility-based regime labeling to assess stability during turbulent episodes. Quantile LSTM is benchmarked against standard VaR baselines (Historical Simulation, parametric normal, EWMA, and GARCH(1,1) with normal and Student-t innovations) using violation rates, Kupiec and Christoffersen backtests, pinball loss, and loss-based model comparison. Results indicate good aggregate calibration for Quantile LSTM, with violation rates of 0.554% (6/1083) for VaR 1% and 5.079% (55/1083) for VaR 5%, passing unconditional coverage. However, at there are indications of clustered exceedances, highlighting the value of rolling and regime-based diagnostics. At the 1% tail, GARCH-t achieves lower and significantly different pinball losses, reinforcing heavy-tail models as strong benchmarks for extreme risk. Beyond accuracy metrics, the regime-sensitive diagnostics provide actionable signals for deploying Quantile LSTM within automated risk-monitoring pipelines, where rolling exceedance alarms and adaptive retraining can improve operational reliability under volatility regime shifts.]]>
