基於多種變數選擇技術與極限學習機於飯店業銷售預測

HOTEL SALES FORECASTING BASED ON VARIABLE SELECTION TECHNIQUES AND EXTREME LEARNING MACHINE

李天行(Tian-Shyug Lee) ； 呂奇傑(Chi-Jie Lu) ； 鄭婷方(Ting-Fang Cheng)

《品質學報》 24卷6期 (2017/12) Pp. 411-430

https://doi.org/10.6220/joq.2017.24(6).02

摘要

由於飯店業的商品具不可儲存且供給彈性小的特性，讓如何建構有效的銷售預測模式變成為重要的議題。且因飯店業的銷售容易受許多因素的影響，找出重要的銷售預測變數提供企業進行決策時之參考，可提升銷售管理效能。本研究結合三種不同特性的變數選擇方法─逐步迴歸 (stepwise regression)、基因演算法 (genetic algorithm)、貪婪演算法 (greedy algorithm)，與新興的機器學習預測技術─極限學習機 (extreme learning machine, ELM) 提出三個混合式銷售預測模式。且不同於以往的研究，在候選預測變數上，除了使用總體經濟指標外，也根據股票技術分析的概念，利用歷史銷售資料產生技術分析變數作為預測變數。本研究以晶華飯店、國賓飯店與亞都飯店等三家臺灣上市飯店業的月銷售額為實證對象。實證結果顯示，三個混合模式的預測績效均優於單純ELM模式，並且greddy + ELM模式是適用於晶華飯店的銷售預測模式，國賓飯店與亞都飯店則是使用stepwise + ELM模式能產生最佳的預測結果。且對晶華飯店與亞都飯店而言，最重要的預測變數為「去年同期銷售額」，代表這兩家飯店的銷售具有穩定性，季節性的影響明顯。而國賓飯店的重要預測變數則為「MA(3)」，代表其銷售較受短期因素的影響。

關鍵字

飯店業 ； 銷售預測 ； 變數選擇 ； 極限學習機

並列摘要

Due to the characteristics of hotel product are fixed capacity, high fixed costs and perishable inventories, sales forecasting has therefore always crucial for hotel sales management. When constructing a sales forecasting model, discussing and understanding the important predictor variables can help focus on improving sales management efficacy. Aiming at to select appropriate predictor variable and construct effective forecasting model, this study combines three variable selection methods, including stepwise regression, genetic algorithm and greedy algorithm, and extreme learning machine (ELM) to construct hybrid sales forecasting models for hotel industry. For discussion of the important predictor variables for sales of hotels, the macroeconomic indicators and technical indicators of historical sales data are used in this study as candidate predictor variables. The monthly sales data of the three listed international hotels including Regent Hotels & Resorts, Ambassador Hotel and Landis Hotels & Resorts are used as the illustrative examples. The experimental results indicate that the proposed hybrid sales forecasting models outperform the single ELM model. The most suitable model for forecasting sales of Regent hotel is the greedy + ELM model. For Ambassador and Landis hotels, the most proper model is the stepwise + ELM model. Furthermore, the ＂Last Year Sales＂ variable is the most important predictor variable for Regent and Landis hotels. The ＂MA(3)＂ variable is the most influential variable for Ambassador hotel.

並列關鍵字

hotel industry ； sales forecasting ； variable selection ； extreme learning machine

參考文獻

Chen, K. Y.,Wang, C. H.(2007).Support vector regression with genetic algorithms in forecasting tourism demand.Tourism Management.28(1),215-226.

Chen, M. H.(2010).The economy, tourism growth and corporate performance in the Taiwanese hotel industry.Tourism Management.31(5),665-675.

Girlich, E.,Kovalev, M. M.,Vasilkov, D. M.(1997).Greedy sets and related problems.European Journal of Operational Research.101(1),74-80.

Goh, C.,Law, R.(2011).The methodological progress of tourism demand forecasting: a review of related literature.Journal of Travel & Tourism Marketing.28(3),296-317.

Han, K. S.,Viau, A. A.,Kim, Y. S.,Roujean, J. L.(2005).Statistical estimate of the hourly near‐surface air humidity in eastern Canada in merging NOAA/AVHRR and GOES/IMAGER observations.International Journal of Remote Sensing.26(21),4763-4784.