Demand estimation via machine learning

Abstract

In the TV home shopping, there are two main problems to construct broadcasting timetable to maximize profit for the traditional merchandisers. The first is that they cannot predict the sales amount for product on less broadcasted time easily, and the second is that they are hard to use the large number of explain variables that could be related on the demand for the product until they do not use any sophisticated model. To overcome these problems, we construct the demand forecasting model on TV home shopping context using statistical model and machine learning model and compare the performance between them. In the result, the demand forecasting performance on products during broadcasting is higher when the model is constructed by a tree-based machine learning algorithm compared to when it is constructed by a traditional statistical model. Specifically, the light gradient boost model has the best performance, and the ordinary least squares model has the worst performance. In particular, we determined that, for calculating mean squared error to measure performance, the error of low-true value is more underestimated than the error of high-true value, so we suggest the weighted mean squared error, which is weighted more on low-true value and less on high-true value.