Predictive Modelling of Sales Volume for Fast Moving Consumer Goods

This work is devoted to accurate analysis of predictive models of time-series in terms of limited information in the are of retail sales. In this work, different approaches to forecasting were compared: classical one that relies on econometric approach and another one with the use of machine-learning modelling. The main hypothesis of the research was that models in machine-learning approach in terms of the absence of data of dependent variable with the lag of one year are able to perform relevant results, while econometric approach can beat competitive score only having the access to data of previous periods. For the research, the data of Rossman sales were used. The best result was performed by the model of Gradient Boosting after the fine-tuning of hyper-parameters with the help of cross-validation technics. The models of decision trees, random forest, adaptive weighted lasso regression and autoregressive model with seasonality trend are also presented in the study. The models of autoregression and adaptive weighted lasso performed the score three times worse than machine-learning-based models, however, it is worth of attention that these models had a significant accuracy gain conversely to machine-learning ones. The metric for model evaluation was chosen – RMSPE. The aspect that is also worth of attention is the determination of factors that contributed the most to the deviation of sales dynamics. The model of Gradient Boosting, Adaptive Weighted Lasso regression and Random Forest highlighted Competitors’ Distance, Promo and Day of Week. From the business point of view, the value of this research is in the fact that the customers of this chain tend to choose its stores not because of their loyalty but because of the lower prices due to the promotions and inconvenient or far location of competitive chains stores.