Abstract
Online bookings are major inputs to modern local businesses. Accurate forecasts of online booking demand are crucial for service providers to provide better customer services and allocate resources more efficiently. We develop a forecasting procedure for forecasting time series of weekly online restaurant bookings, using data from EZTABLE, the biggest online booking platform in southeast Asia. Because restaurant demand is greatly impacted by holidays and other special occasions, we study different possibilities of capturing such information to achieve accurate forecasts on non-special periods. The literature on forecasting bookings is mostly about hotel reservations, where the time series of bookings for each hotel are considered separately. We develop a method that is useful for platforms that have data on many service providers (such as many restaurants or hotels). In particular, we generate forecasts for each restaurant’s reservations by creating models that use data from other restaurants. We focus on how to identify special weeks so that they can be modeled separately, and use different approaches to solve challenges that arise in forecasting weekly reservations data for many restaurants in the presence of unusual demand on special weeks. These challenges include the weekly frequency of forecasting, complex seasonality, and restaurant strategies. We develop and compare several approaches, including using calendar dates, using historic data from a single restaurant, and learning from multiple restaurants.