簡易預測：單變量時間序列集成模型於銷售預測之應用

天真預測法（Naive Forecasting）、移動平均法（Moving Averaging）、指數平滑法（Exponential Smoothing）等單變量時間序列模型（Univariate Time Series Model）已廣泛地應用在銷售預測（Sales Forecasting）的議題上，為進一步提高模型的預測精度，本研究提出了一種可整合上述模型於一體的單變量時間序列集成模型（Ensemble Model），除了各模型的超參數（Hyper-parameter）之外，用來整合各單變量模型的超參數皆可透過遺傳演算法（Genetic Algorithm, GA）自動估算的，免除人為調整多個超參數的困擾。本研究以一家位於台灣新北市三和夜市中的內衣店為例，以其2017年到2019年的周銷售量數據進行模型的訓練與測試，實驗結果顯示，本研究所提方法能夠為該家商店精進模型的預測精準度，其測試數據的平均絕對誤差百分比（Mean Absolute Percentage Error, MAPE）達0.333，與最佳的傳統方法相比約降低了0.01。此外，以上所有的訓練與測試過程都是在Microsoft EXCEL環境中進行，這使得沒有具備統計知識與商用統計軟體的一般用戶均能輕易地實現這套想法，符合EZ精神。

關鍵字

銷售預測；單變量時間序列模型；集成學習；基因演算法

並列摘要

Univariate time series methods for example the naive forecasting, the moving averaging, the exponential smoothing, etc. were widely used in the sales forecasting. To improve of prediction precision, this research proposed a univariate time series ensemble model which was able to integrate above univariate time series methods. Not only the hyper-parameters but also the fusion weights of each univariate model were determined by the genetic algorithm (GA) where the difficulty of tuning manyhyper-parameters manually was alleviated. The weekly sales volume ranged from 2017 to 2019 for an underwear store at the Sanhe night market in New Taipei city, Taiwan was trained and tested. Experimental results showed that the proposed ensemble model could provide a higher prediction precision. The mean absolute percentage error (MAPE) for testing data was about 0.333 which was about 0.01 lower than the best traditional method. All procedures were conducted by a user-friendly software, Microsoft EXCEL, which made it easier for users to implement the idea without statistical knowledge and packages.