On Adaptive Prediction of Nonstationary and Inconsistent Large Time Series Data

The use of time series prediction results in benefits for an organization. Forecasting efficiency relies on applied prediction formula and quality of data received from technical devices and manually inputted. They are often of low quality, with inconsistencies. However, high data quality is crucial for efficient forecasting/prediction purposes (also event detection from time series and pattern recognition), in particular during large data sets processing (often heterogeneous, including data obtained from IoT devices). Such processing should cover inconsistency analysis, interpolation of missing/lacking data, as well as the use of data pre-transformations. The paper presents problems of inconsistent, nonstationary data prediction on the example of stock level daily forecasting. Selected methods of time series interpolation are outlined. Results of implementation of algorithms for short-term time series prediction are illustrated and discussed. Prediction quality measured based on errors values calculated both in total and in a moving window is discussed. A concept of an adaptive algorithm based on a change in the prognostic formula depending on short-term characteristics of time series is outlined.