The Methodology of Preliminary Expert Assessment of the Quality of Time Series Data for Forecasting Purposes

The article examines the problem of the dependence of the accuracy of forecasts in complex systems on the quality of the source data. The key aspects determining the suitability of historical data for forecasting purposes are analyzed: the forecasting horizon, the volume and representativeness of the sample, consistency, stationarity, the presence of anomalous observations, relevance, frequency and regularity of data receipt. Special attention is paid to the methodological limitations of classical approaches to data quality assessment. An original method of preliminary expert assessment is proposed, based on the principles of system analysis and including three successive stages: 1) assessment of the sufficiency of the data volume using the modified Pareto principle, 2) verification of the regularity of data through the coefficient of variation, 3) identification of anomalies using the “forecast tube” concept. To quantify the importance of factors, a weighting system has been developed based on the criteria of decision theory (Laplace, Wald, Savage, Hurwitz). The methodology makes it possible to systematize the process of preliminary analysis, minimize the risks of using unrepresentative data, and reasonably choose forecasting methods. The practical significance lies in reducing the time spent on data preparation and increasing the reliability of predictive models in conditions of uncertainty.