Long short-term memory network based deep transfer learning approach for sales forecasting

Abstract

The general flow chart of the proposed approach is shown in Figure A.Purpose: The aim of this study is to increase the forecasting accuracy and reduce the computational cost of the deep learning models for sales forecasting. For this purpose, a long short-term memory (LSTM) based deep transfer learning approach is proposed.Theory and Methods: Deep transfer learning enables the transfer of the knowledge acquired in a source domain and task to a target domain and task. In the proposed approach, source selection is performed according to the similarities between the source and target sales datasets, and edit distance with real penalty (ERP) is adopted for this purpose. The most similar source dataset is used for training the LSTM network, which allows extracting the temporal dependencies within the dataset. After the parameter transfer, the LSTM network is re-trained with the target dataset. Eventually, the proposed ERP-LSTM-TL model is obtained for sales forecasting. Results: Experiments with various sales datasets showed that transfer learning improved the forecasting accuracy in 38 out of 46 source and target dataset combinations. On the other hand, negative transfer learning was observed in the remaining eight combinations. The proposed ERP-LSTM-TL method prevented the negative transfer in all target datasets. Also, it yielded superior performance compared to the traditional forecasting and machine learning methods, and reduced the training time of the deep learning models.Conclusion: Experimental results showed the effectiveness of ERP-LSTM-TL in sales forecasting for different products and different sectors. Manufacturers, retailers and distributor companies can obtain cost and time savings using the proposed approach.