Study of the influence of changing the parameters of the ARIMA model on the quality of the forecast for short data sets

Time series, i.e. data collected at various times. The data collection segments may differ depending on the task. Time series are used for decision making. Time series analysis allows you to get some result that will determine the format of the decision. Time series analysis was carried out in very ancient times, for example, various calendars became a consequence of the analysis. Later, time series analysis was applied to study and forecast economic, social and other systems. Time series appeared a long time ago. Once upon a time, ancient Babylonian astronomers, studying the position of the stars, discovered the frequency of eclipses, which allowed them to predict their appearance in the future. Later, the analysis of time series, in a similar way, led to the creation of various calendars, for example, harvest calendars. In the future, in addition to natural areas, social and economic ones were added. Aim. Search for classification patterns of time series, allowing to understand whether it is possible to apply the ARIMA model for their short-term (3 counts) forecast. Materials and methods. Special software with ARIMA implementation and all need services is made. We examined 59 data sets with a short length and step equal a year, less than 20 values in the paper. The data was processed using Python libraries: Statsmodels and Pandas. The Dickey - Fuller test was used to determine the stationarity of the series. The stationarity of the time series allows for better forecasting. The Akaike information criterion was used to select the best model. Recommendations for a reasonable selection of parameters for adjusting ARIMA models are obtained. The dependence of the settings on the category of annual data set is shown. Conclusion. After processing the data, four categories (patterns) of year data sets were identified. Depending on the category ranges of parameters were selected for tuning ARIMA models. The suggested ranges will allow to determine the starting parameters for exploring similar datasets. Recommendations for improving the quality of post-forecast and forecast using the ARIMA model by adjusting the settings are given.