| 摘要: | 過去十餘年，全球經濟深受新冠疫情、中美貿易摩擦、烏俄衝突以及整體經濟放緩等多重因素的衝擊，金融市場波動愈發劇烈。在現今動盪環境中，黃金期貨因作為金融市場重要指標與兼具避險與投資屬性而備受矚目。然而，受到市場變化與政治因素的影響，其價格波動劇烈且隨機性與非線性程度高，導致為價格預測帶來極大的跳戰。
在過往的研究中，學者們提出了利用許多深度學習模型來進行期貨價格預測。然而，現有研究仍侷限於時域特徵，對價格序列中頻域形態著墨甚少。 因此，本研究提出一種基於梅爾頻率倒譜係數（MFCC）的特徵提取方法，針對價格的頻域進行分析以獲得更多的特徵資訊，成為一種新的黃金期貨價格預測方法，期望在準確度和有效性方面取得顯著的提升。
本研究透過MFCC提取黃金期貨與相關指標的特徵數據，並將其切割為短期、中期和長期數據，藉此模擬在不同時間顆粒度下的訓練特徵，輸入LSTM模型進行訓練。最終，本研究結果證將MFCC特徵提取技術與LSTM時序學習相結合，可有效提升模型預測準確度，為未來的研究提供了一種新的方法與方向。;Over the past decade, the global economy has been impacted by the COVID-19 pandemic, Sino-U.S. trade frictions, the Russia–Ukraine conflict, and a broad-based economic slowdown, all of which have amplified volatility in financial markets. Under this circumstance, gold futures, which are widely viewed as both a indicator of macro-economic risk and a reliable hedge, have attracted increased attention. Yet, the series of gold futures is highly random and nonlinear, with price swings strongly influenced by geopolitical and macroeconomic shocks, making accurate forecasting particularly challenging.
Prior studies have explored a variety of deep-learning models for futures price prediction, but most approaches rely primarily on time-domain features and devote limited attention to the frequency-domain characteristics embedded in price data. To address this gap, the present work introduces a Mel-Frequency Cepstral Coefficient (MFCC)–based feature-extraction scheme that analyzes gold-futures prices in the frequency domain, enriching the information available to the forecasting model in this study and aiming to improve both accuracy and robustness.
In this study, MFCC features are extracted from gold-futures prices and related market indicators, then segmented into short- term, medium- term, and long-term windows to reflect multiple temporal granularities before being fed into a Long Short-Term Memory (LSTM) network. Empirical results confirm that combining MFCC-derived frequency-domain representations with LSTM sequence learning markedly enhances predictive performance, offering a promising alternative for future research on gold-price forecasting and broader financial time-series analysis. |
