Abstract
Time-series forecasting is a popular technique and is used for different purposes in many fields. However, one big problem exists in many common forecasting algorithms is the short-history problem. Our study presents a 3-step method that can automatically provide forecasts for items with very short histories. This 3-step method composes of time-series clustering, time-series matching, and forecasting methods. There are several different strategies in each step, so 12 combinations are conducted in our study. We compare the forecasting performance of different combinations using real sales data from iCHEF, a Taiwanese platform company. The results show that time-series matching methods influence forecasting performance significantly. Besides, similar historical data is useful information while forecasting for items with short histories.