雞蛋價格分析:時間序列與機器學習模型之應用

Abstract

雞蛋為我國民眾日常飲食中不可或缺之基礎食材，亦為重要的動物性蛋白來源之一。近年來，受 2021 年 COVID-19 疫情、2022 年烏俄戰爭及 2023 年禽流感等重大事件影響，雞蛋市場供需結構出現顯著波動，不僅加重家庭食品支出負擔，亦衝擊蛋農之經營穩定性，因而成為社會高度關注的重要議題。本研究以 2008 年 5 月至 2025 年 6 月之國產雞蛋價格為研究對象，分別採用原始資料與對數轉換資料進行實證分析，並納入多項外生變數及其落遲項，以刻畫雞蛋價格之動態調整過程。研究方法上，同時運用 SARIMAX 時間序列模型與隨機森林模型，探討影響雞蛋價格波動之關鍵因素，並比較不同模型架構下之預測表現與變數重要性。 實證結果顯示，在 SARIMAX 模型中，需求面變數如鴨蛋價格與進口雞蛋數，僅於特定落遲期下呈現顯著影響，顯示其對國產雞蛋價格之替代效果相對有限。相較之下，供給面變數對價格形成具有較為穩定且顯著的影響，其中產蛋箱數與飼料價格皆呈現顯著正向關係，反映供給調整機制與生產成本轉嫁為雞蛋價格形成之關鍵因素；而換羽隻數則呈現顯著負向影響，顯示蛋農會依價格水準調整生產行為。此外，COVID-19 事件虛擬變數於各模型中皆呈現顯著正向效果，顯示疫情期間之供給擾動與成本上升對雞蛋價格具有明確推升作用。隨機森林模型之結果則顯示，落遲雞蛋價格於變數重要性排序中居於首位，顯示雞蛋價格具有高度延續性，而需求與成本相關變數之相對重要性則較低。在原始資料架構下，SARIMAX 模型較能捕捉短期價格波動，但易出現過度預測現象；相較之下，隨機森林模型之預測結果較為平滑，其中隨機森林模型一具有最低整體預測誤差，為原始資料下表現最佳之模型。在對數資料架構下，SARIMAX 模型較能掌握價格變動之趨勢方向，而隨機森林模型在預測誤差控制與穩定性上仍具相對優勢，其中隨機森林模型四為對數資料下預測表現最佳。綜合而言，SARIMAX 模型有助於趨勢判斷與經濟結構之解釋，但對價格轉折之反應相對落遲；隨機森林模型則在整體預測準確性與穩定性方面表現較佳。兩類模型各具優勢，具備良好之互補性，可為雞蛋價格分析與預測提供更為全面之實證依據。

Eggs are an indispensable staple in the daily diet of the population and an important source of animal protein. In recent years, major events—including the COVID-19 pandemic in 2021, the Russia–Ukraine war in 2022, and avian influenza outbreaks in 2023—have caused substantial disruptions to the supply–demand structure of the egg market. These shocks have not only increased household food expenditures but have also undermined the operational stability of egg producers, thereby drawing widespread social attention.This study investigates domestic egg prices over the period from May 2008 to June 2025. Both raw price data and logarithmically transformed data are employed for empirical analysis, with multiple exogenous variables and their lagged terms incorporated to characterize the dynamic adjustment process of egg prices. Methodologically, this study applies both the SARIMAX time-series model and the random forest model to identify key determinants of egg price fluctuations and to compare forecasting performance and variable importance across different model specifications. The empirical results indicate that, within the SARIMAX framework, demand-side variables such as duck egg prices and imported egg quantities exert statistically significant effects only at specific lag lengths, suggesting that their substitution effects on domestic egg prices are relatively limited. In contrast, supply-side variables display more stable and significant influences on price formation. Specifically, the number of egg-laying boxes and feed prices exhibit significant positive relationships with egg prices, reflecting the roles of supply adjustment mechanisms and cost pass-through in price formation, while the number of molting hens shows a significant negative effect, indicating that producers adjust production behavior in response to price levels. In addition, the COVID-19 event dummy variable is significantly positive across all model specifications, demonstrating that supply disruptions and rising production costs during the pandemic exerted upward pressure on egg prices. Results from the random forest models further reveal that lagged egg prices rank highest in terms of variable importance, indicating strong price persistence, whereas demand- and cost-related variables are relatively less influential. Under the raw data specification, the SARIMAX model is more effective in capturing short-term price fluctuations but tends to over-forecast. By contrast, the random forest models generate smoother predictions, with Random Forest Model 1 achieving the lowest overall forecasting error and thus performing best under the raw data framework. Under the logarithmic data specification, the SARIMAX model better captures price trend movements, while the random forest models continue to demonstrate superior performance in error control and stability, with Random Forest Model 4 delivering the best predictive performance.Overall, the SARIMAX model is well suited for trend identification and economic structure interpretation but responds more slowly to price turning points. In contrast, the random forest model performs better in terms of overall forecasting accuracy and stability. These two approaches exhibit complementary strengths and together provide a more comprehensive empirical basis for egg price analysis and forecasting.