Abstract: This paper develops a reproducible framework for predicting gold-price direction and converting the forecasts into a currency-hedged SHFE-COMEX spread-trading strategy. Daily market data from January 2018 to December 2024 are divided chronologically into training, validation, out-of-sample testing, and frozen-model forward paper-trading periods to limit data leakage. A hybrid LSTM-Transformer-Attention model achieves 63.7% directional accuracy in 2023, although paired tests do not establish statistical superiority over the Transformer benchmark at the 5% level. The trading evaluation incorporates next-session execution, commissions, slippage, holiday mismatches, contract rolls, integer position constraints, and an SGX USD/CNH futures hedge. Under the stated assumptions, hedging reduces volatility and maximum drawdown and improves risk-adjusted performance. The findings provide preliminary, reproducible evidence of economic usefulness, but not proof of durable arbitrage profitability without intraday executable quotes and complete trade-level records.

Keywords: Gold futures, Deep learning, LSTM, Transformer, Cross-market arbitrage, USD/CNH hedging, Backtest bias, WSEAS format

Pages: 1755-1761