1996年起台北市政府推動台北年貨大街之活動，一舉將迪化街與年貨大街畫上 等號，每逢歲末年初，上迪化街辦年貨幾乎成了北台灣人的全民運動。據迪化街商 家表示，在年貨大街活動期間各種商品銷售的量，往往是平時每月銷售量業績的數 十倍、甚至百倍，驟升的業績帶來興奮的喜悅，但也伴隨著沉重成本的壓力，貨品 若無法在短短15至20天的活動期間內販售完畢，平時銷售量可能因無法攤銷以導致 高成本的積壓。因此，要如何透過有限的資訊預測貨品的準備量乃本研究所欲探知 的主題。 傳統預測方法如天真預測法與移動平均法都僅透過歷史資料對未來做預測，往 往會忽略其他的影響因子；而相對較可靠的時間序列預測法、線性迴歸分析以及類 神經網路等方法卻只在累積數據較多的情形下才可行；灰預測(Grey Forecasting)除了 簡單易懂外，更能在少量數據的情境下進行預測，本研究透過NGM(1, N)之非線性多 變量灰預測模型進行花生銷售量之預測，以公開或私人且少量的多變量數據來預測 下一年度之銷售量，以作為進貨量的參考。此外，研究中亦引入基因演算法(Genetic Algorithm)對NGM(1, N)進行超參數微調(Hyper-parameter Tuning)與變數選擇(Feature Selection)，以減少模式執行時的人為干擾。

In 1996, the "Taipei New Year's Grocery Street", which was promoted by the Taipei City Government, had combine Dihua street with the “Taipei New Year Market (TNYM)” together. When to time has come to the end of year, citizens whom live in north part of Taiwan will visit Dihua street and prepare groceries for the New Year. Groceries, which had traded during the TNYM, had created more than ten times of profit. Sellers said, “Happiness and pressure were two feelings that brought to sellers during this event.” The quantity of commodities, the cost of commodities, and the time were major problems for sellers need to be concern. So, for sellers to predict the number of commodities before the event, which only had short period, well be the major subject in this essay.Traditional predictive methods such as Naive Forecasting and Moving Averages are all method for extend predictors to the historical data, but those methods will neglect the influence factors that are not mention in historical data. Time Series Forecasting, Multiple Regression Analysis and Artificial Neural Network are more reliable methods that only can be use and comply in larger dataset. The grey multivariable model is a simple and easy to understand method, it is the only method that requires small and incomplete information to do the prediction. This study will use nonlinear grey multivariable model, NGM(1, N) with small amount and readily available multi-variable data to forecast the next period’s sales volume, which will reference for the purchase volume for the event. Besides, this study leases several adjustable hyper-parameters such as the smoothing factor, the initial condition, and power exponents as well as the feature filter and then integrates all of them into the model. Furthermore, the genetic algorithm (GA) is introduced to alleviate the problem of manual selection of those hyper-parameters and features.