Model setting using stationarity criteria for time series forecasting

The article discusses the possibility of using the information on the stationarity of residuals to improve the procedure of forecasting nonstationary time series. In the traditional approach, this procedure is used only to confirm or reject the hypothesis of nonstationarity of residuals. In this article, the stationarity test is used for fine-tuning of hyperparameters of the forecasting models. The technique is based on the Granger cointegration approach property to find a statistically significant relationship between time series. The author used the p-value of stationarity tests as a loss function. Economic and generated time series were used as data for verification. The experiments have shown that this approach is often more effective in comparison with the traditional methods of tuning models.