Quantitative research shows that short look-back windows (1-7 days) often outperform longer histories when training LSTM and GRU models for next-day forex direction prediction. A 2025 study across eight USD pairs found that freely floating pairs like USD/JPY favor the shortest context, while policy-managed pairs like USD/CNY can benefit from moderately longer windows. This suggests market efficiency varies by pair, and model architecture should be tuned per instrument rather than applied uniformly.