Predictive Modelling for the Dynamics of Consumer Activity by Methods of Machine Learning and Big Data of Payment Systems

The article contains a feasibility analysis of using big data of payment systems for building forecasting models for the dynamics of consumer demand as a core factor of the retail turnover. The relevance of this research stems from the need for prompt and reliable estimates of economic activity in a highly uncertain and volatile economic environment, especially due to the large-scale Russian invasion of Ukraine in 2022. It is demonstrated that using big data as an alternative information source is an expedient option to minimize time lags intrinsic in the official statistics, and to enhance the accuracy of assessment of the current economic performance. A set of relative indicators was built to measure various operative aspects of the payment infrastructure and consumer behavior: ratio of payment card activity, availability of POS terminals, intensity of using payment infrastructure, and proliferation of contactless technologies. Principal component analysis was applied to this set of indicators to build Integral index of consumer activity (IICA), summing up multidimensional information and preserving a larger part of variation in output data. It was found that the first principal component explained more than 80% of the overall variance. The results show that IICA has a high information capacity and can be interpreted as a leading indicator of change in the consumer demand. To estimate IICA impact on the dynamics of consumer demand, an econometric model was built by least squares with account for lag dependencies. Index of Physical Volume of Retail Turnover was used as an output variable, interpreted as a proxy indicator for the effective consumer demand. The results of the modelling show statistical significance of IICA, and a high explanatory capacity of the model, thus confirming the existence of a stable correlation between payment activity indicators and macroeconomic dynamics. The research includes a comparative analysis of the efficiency of classical econometric methods and advanced methods of machine learning, Prophet model of time series in particular. The model demonstrated high-quality approximation on a training sample but was found inefficient for forecasting under sudden economic changes. The results showed that the above methods could not ensure quality of forecasting compared with the basic econometric model. Models of machine learning demonstrated low capacity for summarization on test samples, which might be attributed to a limited data volume and the existence of structural breaks in time series. It is concluded that a core factor enhancing the accuracy of forecasting is not so much the model’s complexity as the quality and information capacity of input data. Using big data of payment systems allows one to build leading indicators measuring more adequately the current economic performance and ensuring a prompt response on changes in the consumer behavior. The practical significance of this research is that the proposed approach is applicable for monitoring and forecasting consumer demand in the mode approximated to real time, which is especially important for managerial decision-making in the conditions of crises.