An Empirical Evaluation of BiLSTM-Attention for Chinese Soybean Futures Price Forecasting

As agricultural commodities become increasingly financialised, reliable soybean futures forecasts are important for risk management and market monitoring. This study examines the main No. 1 soybean futures contract listed on the Dalian Commodity Exchange using a multi-source daily dataset spanning from February 2015 to December 2025. Ten predictors cover domestic price-volume information, cross-market transmission, the macro-financial environment, and supply–demand fundamentals. To address potential look-ahead bias, the forecasting protocol is strictly causal: monthly fundamentals are lagged by one month before daily alignment; Chicago Board of Trade (CBOT) settlement information is matched only from dates strictly preceding the corresponding Dalian Commodity Exchange (DCE) date; Min-Max scalers are fitted on the training sample only; and wavelet denoising, when used, is performed separately within each 60-trading-day historical input window. Models are trained on target dates through 2022, validated on 2023–2024, and evaluated on the fully held-out 2025 sample. Five fixed random seeds, a random-walk benchmark, directional accuracy, and Diebold–Mariano (DM) tests are used to assess robustness. The results show that thediscrete wavelet transform–bidirectional long short-term memory with attention (DWT-BiLSTM-Attention) specification is sensitive to random initialisation and has significantly higher squared forecast loss than the random walk for the five-seed mean forecast. Among the neural-network specifications considered, the DWT-LSTM model provides the lowest mean root mean square error (RMSE) and the smallest RMSE dispersion across seeds, although its predictive accuracy is not statistically different from the random walk. The ablation results further show that causal DWT and the ten-feature specification do not deliver robust incremental gains, while the incremental effects of attention and bidirectionality are not statistically significant. These findings highlight the importance of leakage-controlled preprocessing, strong naive benchmarks, and repeated-seed evaluation in financial time-series forecasting.