The Research On Forecast Of Hotel's Online Sales Based On The Ensemble Method

With the prosperous development of the mobile Internet,the way of ordering hotel online by mobile phone and other devices has become increasing popular among consumers.It is the core of revenue management and platform management to predict the online sales of hotels in the future by using the massive data accumulated on the order platform.However,the hotel's sales data are usually influenced by complex market factors,which are characterized by complicated fluctuations and obvious holiday effects.So,the traditional single model is difficult to accurately capture the fluctuation characteristics of hotel sales.What's more,the portability of a single model is always poor.To solve above problems,this paper proposes an algorithm based on ensemble method to predict the hotel's online sales.The algorithm consists of two models,one is the hotel's online sales forecast model during the non-holiday period based on residual fitting,the other is the hotel's online sales forecast model during the holiday period based on Lasso.Before prediction,the algorithm firstly determines the period of the predicted time and the chooses the corresponding model to predict.For the hotel's sales forecast model during the non-holiday period,this thesis designs and realizes an ensemble method based on the residual fitting according to the characteristics of strong periodicity and randomness of hotel's online sales.This method first uses FBProphet model and LightGBM model to predict the rules and random items of hotel's sales respectively.Then,the results of the two models are combined to obtain the final prediction results.For the hotel's sales forecast model during the holiday period,according to the fact that sales data has the characteristics of explosive,the correlation between the date,advanced booking volume,the historical sales quantity and the hotel's online sales is analyzed.Then,the feature engineering and Lasso algorithm are built for modeling and forecasting.In addition,this paper also puts forward the detection method of “abnormal hotel” to detect the hotel which may appear the accidental sales anomalies.Some corrective strategies are further designed to get the final prediction.The algorithm proposed in this thesis solves the problem of poor portability of single model in hotel's online sales prediction scene.What's more,the algorithm also includes the detection and correction method of the abnormal hotel,which solves the problem that the model can not predict the reliable result of the abnormal hotel.This paper adopts the real sales data of a well-known domestic ordering platform as experimental data,and proves that the performance of the proposed algorithm is better than the single model of LightGBM and FBProphet during the holiday period and the non-holiday period through a series of experiments.