Abstract:

Gold， possessing the dual characteristics of a commodity and a monetary asset， has assumed an increasingly pivotal role in global financial markets. Its price dynamics are profoundly influenced by a wide array of macroeconomic indicators and geopolitical developments， rendering the gold futures time series highly volatile， nonlinear， and non-stationary. These intrinsic properties present significant challenges for accurate short-term forecasting. In response to these challenges， an advanced deep learning framework is proposed， termed KANsLTformer， aimed at effectively capturing the intricate temporal dynamics and multi-scale structures inherent in gold futures price series. The objective is to enhance both prediction accuracy and model robustness in the face of complex market behaviors. At the core of KANsLTformer lies a novel Temporal Convolutional Gated Linear Unit （TCGLU）， which combines temporal convolution with a gating mechanism to selectively extract and emphasize relevant short-term features. This design enhances the model's capacity to focus on local temporal patterns crucial for high-frequency market fluctuations. Building upon this foundation， the framework integrates a customized Transformer architecture that learns global dependencies and dynamic attention patterns across time， enabling the model to capture long-range interactions effectively. Furthermore， the inclusion of the Kolmogorov-Arnold Networks （KANs） module facilitates the modeling of nonlinear feature interactions through KANs， while Long Short-Term Memory （LSTM） units are employed to retain and utilize long-term historical information， further enriching the temporal representation. Together， these components form a synergistic hybrid architecture capable of modeling both high-frequency and low-frequency patterns and adapting to abrupt market changes. The model is empirically validated using real-world gold futures data， encompassing core trading variables such as open， high， low， and close prices， trading volume， and open interest. The evaluation framework encompasses diverse market conditions， including both stable and highly volatile regimes， and employs time-series-specific validation techniques along with robustness assessments under artificially introduced noise. Comparative experiments benchmark KANsLTformer against several comparative models， including ARIMA-GARCH， XGBOOST， LSTM， and CNN-Transformer. The proposed approach significantly outperforms baseline models， reducing Mean Absolute Error （MAE）， Symmetric Mean Absolute （SMAPE）， Root Mean Square Error （RMSE）， and Median Absolute Error （MedAE）. Compared with its strongest ablation variant， it lowers MAE by approximately 43.5% and achieves a Directional Accuracy （DA） of 63.69%， demonstrating the contribution of each module. The superiority of KANsLTformer is further validated through Diebold-Mariano （DM） tests， which confirm the statistical significance of its predictive improvements over competing models at the p<0.001 level under both noise-free and noisy conditions. Notably， under noisy conditions， the model maintains a DA of 58.03%， consistently outperforming benchmarks with statistically significant DM test values. Ablation studies underscore the indispensable contributions of the TCGLU and LSTM components in capturing short-term volatility and long-term dependencies， respectively. The complete configuration of KANsLTformer demonstrates superior stability， robustness， and generalization performance， particularly in volatile market scenarios. In conclusion， a robust and generalizable framework is introduced for financial time series forecasting， offering practical implications for investors， financial analysts， and policymakers engaged in gold futures trading. Beyond its application to the gold market， the methodological advancements presented herein contribute to the broader domain of deep learning-based time series modeling， opening new avenues for future research in financial prediction and decision support systems.

Key words: gold futures price prediction, transformer, KANs, LSTM, TCGLU