國產大宗蔬菜價格預測模型之比較

Abstract

臺灣全年都可以生產蔬菜，但因生產具季節性，加上受天候因素影響，所以蔬菜價格波動明顯，往往有「菜土菜金」的現象發生。目前國內生產超過百種以上蔬菜，又以花椰菜、青花菜、甘藍與結球白菜產量有最大的比重，被稱為大宗蔬菜，在市場上有舉足輕重的地位，故本研究將以這些蔬菜作為研究對象，分析其價格。 本研究利用2012年至2022年間，果菜批發市場價格作為研究標的，並利用傳統統計模型與類神經網絡等二大模型，包含差分整合自迴歸模型(ARIMA)、指數平滑法(ETS)、時間延遲類神經網絡(TDNN)與長短期記憶細胞類神經網絡(LSTM)，對大宗蔬菜價格進行模型配適擬合與估計預測。 研究結果顯示，四種蔬菜價格分析適用的模型不盡相同。綜合而言，花椰菜價格預測以使用與前5期有自相關之ARIMA模型表現最佳，具有相對最小均方根誤差(RMSE)、平均絕對離差(MAD)與平均絕對誤差率(MAPE)以及最高之方向預測正確率；青花菜價格預測以輸入為前1期價格之LSTM模型相對優，RMSE與MAD為四者模型中最小；而甘藍價格的預測，以輸入為前1期的LSTM模型的表現相對突出，RMSE與MAD最小，並約有6成的方向預測正確；至於結球白菜價格的預測，運用與前2期有自相關之ARIMA模型相對其他模型來的良好，MAD、MAPE為四者模型中相對最小。大宗蔬菜最適預測模型所利用過去期數與最快生長期長短也有關聯，耗時較短時間種植的蔬菜最適模型預測需要的期數較短，反之亦然。然而，在實際價格變化過大時，模型預測往往會失準，在實際價格變化小時，LSTM模型預測較佳，而實際價格變動動大時，則是TDNN模型有較好的表現，較不易受到實際價格變化所造成預測效果之扭曲。

In Taiwan, vegetables can be produced year-round. But due to seasonality of production and impact of weather, vegetable prices change significantly, leading to a phenomenon where vegetable prices can suddenly rise like gold or fall like dirt. There are over a hundred types of vegetables produced in Taiwan. Among them, cauliflower, broccoli, kale, and cabbage hold a substantial share in production, being considered major vegetables in Taiwan with significant market influence. Therefore, this study will use these vegetables as research target to analyze their prices. In this study, we use wholesale market prices of from 2012 to 2022 as the research target, and two models: traditional statistical models and neural networks, including the Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing (ETS), Time Delay Neural Network (TDNN), and Long Short-Term Memory (LSTM) neural networks. These models are used to fit and predict the prices of major vegetables. The result shows that the appropriate models for each vegetable are different. In general, for cauliflower price forecasting, the ARIMA performs best using 5-lag price autocorrelation, with relatively minimal Root Mean Square Error (RMSE), Mean Absolute Deviation (MAD), Mean Absolute Percentage Error (MAPE), and the highest direction forecasting accuracy. As for broccoli price forecasting, the LSTM with previous 1-period price as input demonstrates superiority, with the smallest RMSE and MAD. In the case of kale price forecasting, the performance of the LSTM with previous 1-period price as input is relatively prominent, with the smallest RMSE and MAD, and approximately 60% direction prediction accuracy. Regarding cabbage price forecasting, the ARIMA using 2-lag price performs well compared to the other models, with the smallest MAD and MAPE. The length of input used in the optimal models are related to the length of their growth period. Vegetables with shorter cultivation period require a greater number of periods for forecasting, and vice versa. However, when actual price changes largely, model forecasting accuracy tends to deviate. The LSTM performs better when change is small, while for large price change, the TDNN forecasting is less distorted.