The gradient boosting method modification to forecast demand for individual types of goods

The subject area of retail sales is characterized by the presence of many different noises and anomalies in the data. The use of machine learning methods in forecasting such data leads to a decrease in the accuracy of forecasts and the occurrence of the overfitting effect. Despite the adjustment of the existing parameters in the existing gradient boosting algorithms, this problem remains and there is a need for additional improvements to build more accurate forecasts. Purpose of the study: improving the efficiency of the gradient boosting method of decision trees for forecasting product sales. Methods: a method for modifying the gradient boosting is proposed, which uses gradient calculation adjustments using anomaly estimates of the dataset samples by the isolation forest method. Also, to improve accuracy, additional weighting of these estimates using a sigmoidal function with additional arguments to adjust the influence of the weights is proposed. Results: standard approaches to regularisation and data anomaly mitigation are investigated using the XGBoost algorithm as an example. Additional modifications were made to the proposed method using time series clustering and cross-validation. All modifications were tested in various combinations against standard approaches on several datasets and their contribution to the final model was evaluated. Thus, a pattern of increased accuracy of the final model associated with the need to apply clustering for data with heterogeneous groups of goods, as well as a pattern of increased accuracy when trained on cross-validation using the developed method of anomaly weighting was revealed. The pattern of decreasing accuracy when applying the standard approach is noted. Practical relevance: the modified approach of gradient correction by weighting data samples can be used both directly in the specified subject area and in any other where it is necessary to increase the accuracy of the gradient boosting method on anomalous and noisy data. Discussion: further research on gradient boosting modification could be aimed at assessing the impact of redundant features in sales data that lead to overfitting.