| Abstract: | 由於激烈競爭, 消費性包裝產品(CPG)生產業者多利用降價及各種非價格促銷活動來維繫競爭力。然而由促銷所造成的銷售量劇烈變動，以及特定促銷方式與農曆節日等稀疏的觀察資料, 常讓統計型方法捉襟見肘，即使統計型方法一般的預測績效相當可靠，它們大多無法利用背景資訊進行預測。相對的，判斷型預測由於可以利用背景資訊, 因而在業界廣為流行，卻有諸多偏誤與不一致的問題。分解式迴歸模型似乎成為整合上述兩種方法的自然選擇，以其可以納入背景因子如促銷與節日效應於模型，以充份利用背景資訊, 並以「化整為零，各個擊破」的方式有效舒解預測人員的資訊超載負擔。可是這些背景因子常是特殊事件，它們的歷史資料稀少資訊量不足，導致共線性問題的滋生,重要變數被移除, 並讓最小平方的參數估計失真。在本論文的第一階段, 使用了由知識引導的調適型基因演算法 (KGAGA)來估計參數,在目標函數使用MAPE而非一般常用的MSE來評估參數的可能解, 以減輕逸出值的平方計算所造成的不良影響，並依據專業知識設定參數限制範圍, 使所求得的參數更有意義並且更加真實。 值得一提的是, 本論文提出一種有效監測與跳脫區域解陷阱的機制 (DEMA),以最近l期最佳解目標函數值改善的移動平均數 (MAFI(l)) ,偵測區域陷阱, 並以廣域搜尋突變運算子與近域搜尋突變運算子組成的迴圈, 依特定機率在個體解染色體多個位置執行位元對調,以快速增進族群的差異性,跳脫區域陷阱, 同時藉由近域搜尋突變運算子的極低突變率與高雜交率以維繫搜尋的收斂能力, 搭配正常搜尋達成再次收斂。父代與母代隨機選自目標函數值不同的族群, 可紓解部份”選擇最適”所衍生的族群差異快速降低的困擾。KGAGA 的求解速度快於普通GA, 所求得個體解的品質也穩定優於普通GA.倘若預測期有任何可預期的變動未能由模型有效處理，本論文的第二階段提出一個機械式調整機制,由一組額外的方程式, 包括年節前後季節指數的重新定位、比例調整及綜合調整等三種調整方式負責處理。此外也將週末效應列入考量， 穩定而有效精算出相關背景因子在預期的時間偏移後, 節日效應的重新定位以及預期變動後促銷與假日等的混合效果，不涉及人為主觀判斷, 以消除判斷式調整常見的偏誤與不一致現象。針對國內一家CPG品牌業者通路商的實證研究顯示, 本論文使用的KGAGA在面對背景因子時會是較OLS及OGA更好的模型擬合選擇，不管在參數估計及多期樣本外預測都有更好表現。KGAGA 在樣本外多期預測的表現也優於ARIMA, 指數平滑, 及天真法。此外也證實本論文使用的機械式調整機制能有效降低各種淡旺季週預測之MAPE，而主要的改善來自大型調整。
Due to severe competition, price discount and non-price promotion campaigns are very commonly used by consumer packaged goods (CPG) manufacturer to retain competitiveness. However, dramatic change in sales volume in promotion session and sparse observations like specific promotion and lunar holidays’ sales in the dataset make statistical models struggle, even though in general the latter are more reliable in forecasting performance, most of them can’t use contextual information which is exploited in judgmental forecasting, a very common practice in CPG industry, however, the latter is subject to various kinds of biases and inconsistency.Decomposition regression seems to be the natural option to integrate both methods in that it can incorporate contextual factors like promotion effects and holiday effects into the model, by “divide and conquer” it is capable of alleviating the information processing overload of forecasters. However, because these contextual factors usually are special events, related historical data are sporadic and don’t have substantial variations, causing collinearity to arise, it becomes difficult for least square estimators to have a proper estimation against parameters. In this thesis, a domain knowledge guided genetic algorithm (KGAGA) in the first stage is employed to address this issue by using MAPE as fitness function, instead of more commonly used MSE, to evaluate each candidate solution and alleviate the impact of square operations of outliers. Besides, a set of parameter constraints are set up based on contextual knowledge to ensure these parameters derived are truly meaningful and reflective to the real world. In particular, a detect and escape mutation algorithm (DEMA) is employed to detect any local pitfall (suboptimum) with a simple moving average metric, thereafter a loop of combination of broad search with ratio and deep search mutation operators to dramatically increase population diversity until the pitfall has been escaped. The crossover operator in which each parent to mate is selected randomly from a different group of different fitness may partially solve the dilemma of selection of fittest which is the ultimate cause of premature convergence. KGAGA competes favorably with ordinary GA (OGA) in efficiency and significantly outperforms the latter in effectiveness with its multiple-reconverging capability.In case there are anticipated variations in the forecasting horizon which can’t be handled by the regression model alone, a mechanical adjusting mechanism, formulated in a set of supplemental equations encompassing lunar-new-year seasonal index realignment, proportional adjustment of mixed effect of promotions and holidays at forecasting horizon in the second stage, coupled with the consideration of the weekend effect, can be used to deal with anticipated time shifting problem and reassess mixed effect of these contextual factors consistently and effectively without subjective judgments in judgmental adjustment to avoid possible bias and inconsistency therein. Empirical results of a channel retailer of a CPG brand manufacturer in Taiwan reveal that KGAGA can be a better alternative than least square estimators in parameter estimation and multi-period out-of-sample forecasting considering contextual factors. It also beats OGA, ARIMA, exponential smoothing, and NA?VE model in multi-step out-of-sample forecasting. Besides, the proposed mechanical adjustment mechanism could significantly reduce MAPE of weekly forecasts across seasons with improvement mainly coming from large size combined adjustments. |
