Sales Forecasting Based On Combination Model

With the rapid development of Internet, the traditional industry experiencing tremendous pressure and challenges, offline retailing by continuously the impact of e-commerce, the traditional offline store consumption pattern is gradually weakened. The company urgently need an effective way to predict the future sales so that it can provide reliable support for the decision-making.As the huge development of data-mining technology, some kinds of machine learning methods has been improved successfully in sales forecasting. This paper focus on Rossmann-The Germany’s third largest commodity chain store, and try to find a better model to solve the sales forecasting problem. According to the diversity and complexity of the problem, this paper analysis the characteristics of the data in detail based on the Angle of business, and found the factors of the problem mainly for non-linear factors, and it contains the following features:the fault of data in timeline, periodic changes is not obvious, many features has a significant impact on sales, large difference between the features, etc. According to the result of features analysis, we figure out the defects and deficiencies on mainstream sales forecast method. In order to obtain better results, this paper has designed tow models, one is based on random forest and the other is based on GBRT. First we obtain the high quality of the training data through the data preprocessing and feature extraction, then we process tuning parameters during the experiment to reach the lower error results, so we verify the validity of the models. After the experiment we found the GBRT models has a slow training speed with the condition of high prediction precision, so we use random forest with less fitting to initialize the residual error of GBRT, experiments show that the fusion method can effectively improve the training speed of the model and reach a better result than GBRT itself.