In this project, we collaborate with AsiaYo, an online B&B booking platform company headquartered in Taiwan, to work together on solving their business problem by using forecasting methods. One challenge facing AsiaYo is the revenue lost when they are lack of available rooms on holidays or special peak periods. Considering the enterprise level and resource, we find that it will be more affordable and understandable to focus our solution of this business problem on certain popular areas.
By forecasting the room occupancy of popular cities for next month, the operation team in AsiaYo can better prepare for the upcoming demand and have a reasonable reference for making relevant decisions. In the data processing steps, we first aggregate the daily booking order data into daily room occupancy data. To give a more accurate result, we then use the “cumulative number of orders before 15, 30 and 60 days”, “Holidays”, and “Sakura Season” as our external data in our model. We tried out different models such as exponential smoothing, ARIMA, linear regression, and neural networks. Eventually, we decide to use linear regression with external data as our main model by reason of better errors and simplicity. For the benchmark, we use yearly seasonal naive and monthly seasonal naive to compare the performances with our linear model.
We find that there is a big improvement on the accuracy of our result by adding the external data, especially when using the “cumulative number of orders before 15, 30, and 60 days” data. Another interesting point we find is that our results perform better in Japan comparing to Taiwan due to users’ different order behaviors. However, the performance on holidays still needs to be improved. For future work, we will try on modeling only holidays and also bring in other possible related data such as website pageviews or organic search.