Gold Price Forecasting Using a Hybrid CNN–LSTM–Informer Model with Multiscale Inputs

Abstract

Forecasting gold prices is challenging due to their volatility and nonlinear behavior. This paper introduces a hybrid deep learning model combining CNN, LSTM, and the Informer attention mechanism, enhanced by multiscale temporal decomposition and seasonal data splitting. Our approach leverages multiple timeframes—1-hour (1H), 4-hour (4H), and daily (1D) and incorporates technical indicators such as RSI, MACD, Bollinger Bands, and ADX for improved feature representation. On temporally separated test sets from MetaTrader 5 (MT5), the model achieves high R² scores of 0.9988 (1H), 0.9955 (4H), and 0.9853 (1D), outperforming existing methods. To mitigate overfitting, we apply early stopping, temporal validation splits, and strictly prevent future data leakage. While results are promising, further evaluation on unseen markets and longer periods is needed to confirm generalizability. This study sets a robust benchmark for accurate and interpretable gold price forecasting in volatile financial markets.