چكيده به لاتين
Sales forecasting is an important part of business planning and strategy. It is important in the retail industry and the country's production system due to the lack of appropriate methods or strategic models for this phenomenon and the need for research. To arrive at the research model, by reviewing the literature and evaluating machine learning models, it has been shown that tree methods have worked well on real-world data. Sales forecasting improves the performance of businesses and as a result, by providing a strategic model in sales forecasting, can provide inventory and thus increase customer satisfaction. Therefore, the key issue that we will focus on in this research is the most appropriate algorithm among the two algorithms, Light gradient boosting(LGB) and extreme gradient boosting(XGB) to solve the sales forecast problem. The importance of data preprocessing and analysis in sales forecasting with machine learning models is pointed out in this research. In this study, the statistical population is the data of a company in the retail industry that includes 40 stores and one store has been selected as a statistical sample by random and available methods. The data set includes daily sales of goods and to answer the questions, in the first step, the data were pre-processed, which were first identified by the LOF method of discarded data, and then their values were replaced by the KNN method; Then the properties are considered based on the data set and in the next step, the data were given to the recursive model of Light gradient boosting(LGB) and extreme gradient boosting(XGB) and in these two models the Poisson and Tweedie cost function were evaluated. selected data collection from 21/12/2017 to 27/12/2020 data of a store and includes two departments of food and cosmetics. After converting the data into a time series, 19/3/2020 was used as training data, the last seven days were used for testing data, and finally, two selected models were evaluated in two departments with MAE and RMSE. The results showed the LGB in the food department with the Tweedie cost function and the cosmetics department with the Poisson cost function with the mean absolute error and the Root mean squared error of 0.30, 4.8, 0.16 1.47 choose as the optimal algorithm. Consistent with the results obtained, we can expect to achieve a lower error in the sales forecast in the retail business in LGB, versus XGB. In addition, we expect the performance of this technique to be more efficient in the chain stores than in traditional methods.
