網路搜尋強度於嬰兒奶粉進口預測之分析與應用

相信大部分的人都聽過「人類在網路上的行為是最真實的」，然而身為科技巨頭之一的Google則收錄這些真實行為的數據無償供網路使用者取用。這些數據已有許多廣泛的運用，自2006年便有學術領域的研究，一直到這些年來這類研究的議題都還很熱門。台灣是一個天然乳資源相對不充沛的國家，多仰賴進口；以進口需求面的考量，在有嬰兒奶粉需求時，會以英文的關鍵字進行搜尋，因此本研究利用Google Trends Index分別與台灣出生人口、每人總薪資為因子逐步增減變數，並建立多元線性回歸模型(MLR)及自我回歸分佈滯後模型(ARDL)，探討這些因子對於台灣嬰兒奶粉進口額的預測力，並與差分整合移動平均自回歸模型(ARIMA)預測的結果作對照；研究結果的表現以平均絕對百分比誤差(MAPE)、均方誤差 (MSE)、均方根誤差(RMSE)、平均絕對值誤差(MAE)這四項評估指標予以衡量。本研究發現，各個模型的預測結果在不同評估指標下的表現不太一樣；整體三大類模型的表現，有加入Google Trends Index、出生人口數與每人總薪資為解釋變項的MLR與ARDL模型之預測結果，是比沒有加入各解釋變項的ARIMA模型還要好。整體MAPE的結果，MLR模型是優良的範圍，ARIMA模型與ARDL模型是合理的範圍；整體RMSE的結果誤差值是大於MAE，這說明了每期預測的誤差值之間有比較大的振盪。所有的模型相比較，雖然Google Trends Index有助於預測嬰兒奶粉進口額，但是因為網路搜尋活動有許多的行為意函及雜訊的影響，因此在將Google Trends Index移除後，只保留台灣出生人口、每人總薪資為解釋變項的MLR模型，其MAPE預測結果為最優良。

關鍵字

網路搜尋；谷歌趨勢指數；進口需求；時間序列

並列摘要

It’s believed that most people have heard that "the behavior of human beings on the Internet is the most real." Google, however, one of the technology giants, collects these data from Internet, and it’s free for Internet users to use. These data have been widely applied, there has been research in the academic field since 2006, and the topics of such research have been very popular until these years. Taiwan is a country with relatively inadequate natural milk resources and relies heavily on imports. Considering the demand for imports, when there is a need for baby milk, the keywords of Internet search term will be entered by English, so this study uses Google Trends Index, Taiwan’s birth population and the total salary per person as variable factors to predict the import amount of baby milk. The established models are the Multiple Linear Regression Model (MLR), the Autoregressive Distributed Lagged Model (ARDL) and the Autoregressive Integrated Moving Average Model (ARIMA). The performance of predicting results is evaluated by four measurement indicators: Mean Absolute Percentage Error (MAPE), Mean Squared Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). In this study, the performance of these models under these evaluation indicators are more or less different. Overall, the performance of MLR & ARDL model with Google Trends Index, the number of births and the total salary per person is better than that of the ARIMA model. The overall MAPE results, the MLR models are at good level, the ARIMA models and ARDL models are at reasonable level; the overall RMSE value is larger than MAE, which means there is the relatively large oscillation values in each period. Compared with all models, although Google Trends Index is useful to predict the import amount of infant milk powder, due to the influence of many behavior and noises in Internet search activities, the best model is MLR with good level MAPE result, which includes the factor of Taiwanese-born population and the total salary per person but without Google Trends Index.