預測台灣蛋價的模型實作： 時間序列分析與類神經網路

Abstract

雞蛋在臺灣人的日常飲食中扮演著重要的營養供應角色，因此，維持雞蛋價格的穩定性相當重要。本研究從多個來源蒐集2010 年11月1日至2023年4月30日之蛋價資料，缺漏部分以插值法補足。我們蒐集影響蛋價的相關因子，建立三個蛋價預測模型：(1)多元迴歸（Multiple Regression），(2)ARIMAX （Autoregressive Integrated Moving Average with Explanatory Variables Algorithm），(3)長短期記憶（Long Short Short-Term Memory，LSTM）之類神經網路模型。我們以標準的機器學習方法進行訓練和預測，結果顯示，在多元迴歸、ARIMAX和LSTM 模型預測之均方根差（Root-Mean-Square Error，RMSE），分別為9.2412、9.3654和3.0985元。因此，LSTM表現遠遠優於另外兩者，較能捕捉2023年台灣較為異常且漲幅較大的蛋價。本研究提供一個且可行的LSTM 的架構，供後續研究人員或是政府單位建立模型之參考，並提早制定相關措施。

Eggs assume a significant role in the daily dietary intake of the Taiwanese population, emphasizing their crucial nutritional contribution. Consequently, maintaining stable egg prices holds substantial importance. This study amalgamated egg price data from diverse sources spanning November 1, 2010, to April 30, 2023, with interpolated values to address gaps. Pertinent factors affecting egg prices were collected to establish three egg price forecasting models: (1) Multiple Regression, (2) ARIMAX (Autoregressive Integrated Moving Average with Explanatory Variables Algorithm), and (3) Long Short-Term Memory (LSTM) neural network model. Training and prediction were conducted employing standard machine learning methodologies. The outcomes revealed root-mean-square errors (RMSE) for egg price prediction in multiple regression, ARIMAX, and LSTM models as 9.2412, 9.3654, and 3.0985 units, respectively. Consequently, LSTM outperformed the other two models, effectively capturing the pronounced and larger-scale egg price anomalies expected in Taiwan for the year 2023. This research furnishes a practicable LSTM framework, serving as a reference for future researchers or governmental entities aiming to construct models and implement anticipatory measures.