Mamba 模型於黃金價格預測之應用：與 GRU、LSTM 之比較

Abstract

黃金作為全球重要的避險資產，其價格波動受多種經濟與政治因素影響，預測其走勢一直是金融領域的研究重點。本論文針對黃金價格的時間序列預測問題，探討三種深度學習模型——LSTM、GRU 與近期提出的 Mamba——在不同滑動視窗長度下的預測表現。所有模型均以過去 10 至 120 天的資料作為輸入，預測未來 365 個交易日的價格，並以平均絕對百分比誤差（MAPE）作為主要評估指標。 研究結果顯示，LSTM 在多數視窗設定下表現最佳，展現優異的預測能力；Mamba 次之，GRU 在短期視窗下誤差較大。儘管本研究僅針對 Mamba 進行超參數微調，LSTM 仍顯現出最佳的泛化能力。此結果顯示，LSTM 為三種黃金價格預測方法的首選。

As a key global safe-haven asset, gold is influenced by a variety of economic and political factors, making its price forecasting a long-standing focus in the financial domain. This thesis investigates the forecasting of gold prices using three deep learning models—LSTM, GRU, and the recently proposed Mamba—under different sliding window sizes. All models use historical data from the past 10 to 120 days as input to predict the next-day price over the following 365 trading days, with mean absolute percentage error (MAPE) as the evaluation metric. The experimental results show that LSTM consistently achieves the best predictive performance across most window settings, demonstrating high forecasting accuracy. Mamba ranks second, while GRU performs relatively poorly at the shorter ends of sliding window sizes. Despite hyperparameters being set by the Mamba model, LSTM consistently outperforms both Mamba and GRU, except at the shortest sliding window size (10 days), where Mamba has a slight edge. These findings suggest that LSTM is the favored choice for gold price forecasting among the three.