Airfare Forecast Based On Multiple Forecast Models

With the rapid development of civil aviation industry,more and more people choose to travel by air.Before the government deregulated the airline industry,airfare was often determined by air miles.Nowadays,under the condition of free market economy,airlines often consider various factors to set air ticket prices based on their own interests,so as to maximize their revenue.For airlines,airfare forecast technology can help companies make profits.For buyers,airfare forecast technology can predict when the lowest fare will appear,so that buyers can book tickets at the right time.Therefore,airfare forecast technology is of great significance to both supply and demand sides of air tickets.Under the above research background,this paper analyzed the real data from travel websites with statistic approaches,especially the characteristics and influencing factors,and put forward the airfare forecast model for long-term and short-term situations by using relevant theories such as time series and machine learning.First of all,we preprocessed the real data and analyzed the influencing factors.The first step was to obtain a high quality data set through various methods of data preprocessing.Meanwhile,this paper proposed a new price list based on the characteristics of airfare time series.The second step,we explored the influencing factors from internal aspect and external aspect.Using random forest to evaluate the importance of features and select variables,we finally determined the characteristics that affect airfare,such as days left until departure,flight number and day of week.Then,this paper put forward the longterm airfare forecast model and the short-term airfare forecast model respectively.The long-term airfare forecast model was based on the fundamental analysis method.Starting from the influencing factors,we used support vector regression algorithm and random forest algorithm to forecast the fare.This method reflected the interaction between the airfare and the flight information,and was suitable for the airfare forecast with a long period.At the same time,we proposed the PSO-SVR airfare forecast model with particle swarm optimization algorithm to solve the problem of slow convergence speed of parameters.Based on the advantages of parallel combination,we proposed the PSO-SVR-RF forecast model by different weighted combination of SVR and random forest,and compared PSO-SVR-RF forecast model with several popular benchmark models.The experimental results show that the PSOSVR-RF parallel combination model achieves better results in the error evaluation index,indicating that the model has stronger predictive ability and higher accuracy.Finally,this paper proposed a short-term airfare forecast model aimed the situation that the influencing factors of ticket price were difficult to obtain completely.This method simplified the airfare data into a time series of date and price,and only focused on the price without influencing factors,which was more suitable for a short period.According to the data characteristics of airfare,we used time series model to forecast the air ticket price,and used random forest model to adjust the residual value of time series.This paper proposed the SARIMA-RF series combination model and compared SARIMA-RF forecast model with several popular benchmark models.The experimental results show that the SARIMARF series combination model has better fitting effect than the ordinary time series models.