Analyzing the effect of competition in the hospitality industry

Abstract: Maximizing hotel's revenue is a hard and complicated task. A lot of aspects have to be taken into consideration during this procedure. One major variable is the competitors. Identifying hotel's competitors and using their behavior could be a crucial advantage for maximizing hotel's revenue. For this reason, in this thesis the effect of competitors' pricing in the demand forecasting is being studied. Five hotels with approximately 13 months of historical data are available. During the study two GLM models were tested, Poisson and Negative Binomial regression. As a baseline, the models were modeled only with time related features. Competitors' pricing features were added to the baseline models in order to test the effect of the competition. The feature extraction procedure was implemented by the author of the thesis. A LASSO regression was used for feature selection. The effect of competition is tested in terms of the performance of the models and their forecasting behavior. Moreover, it is of great interest to test, whether any of the features or class of features proposed in this study dominates among others. In order for the goals of the project to be accomplished, the evaluation procedure was split in two parts. In the first one, all models were trained into approximately 12 months of historical data. The goodness of fit was tested by a likelihood ratio test and the forecasting behavior for the remaining one month was commented. For the second part of evaluation, a cross validation forecasting error was computed for two different forecasting windows (7 days and 30 days). The evaluation metrics were RMSE, sMAPE and CMAE. The results showed that adding competitors' features improved the goodness of fit of the models. The cross validation forecasting error results differ among hotels. For some of them, adding competitors' features led to an important decrease of the error metrics. For some others, the decrease was not so high, but it still occurred. On average of all hotels, the three metrics were decreased for both 7 and 30 days of forecasting window. For the forecasting behavior, the models with competitors' features were able to capture peaks in specific days, something that the models without competitors could not. The results also showed, that none of the features or group of features proposed in this thesis dominates among others. Consequently, adding competitors' features in the models appeared to have a positive effect to the demand forecasting by improving the performance of the models.