Comparative Analysis of Holt-Winters Algorithms on the Oracle Machine Learning Platform

Abstract

The Oracle Cloud Machine Learning toolkit was used in the work to study time series forecasting methods. The choice of tools was dictated by conducting research at the State Tax University (Irpin) for further use of the results in the work of regional and central divisions of the state tax service, whose information systems are developed on the Oracle platform. Holt-Winters exponential smoothing algorithms, currently represented by the functions of the application programming interface for machine learning models of the PL/SQL package DBMS_DATA_MINING, were investigated. The research used personal income tax data for 2015-2023, obtained from the tax office in the Ivano-Frankivsk region. Higher accuracy (MAE=3,57%) was shown by the algorithm of the Holt-Winters multiplicative exponential smoothing model with a fading multiplicative trend and multiplicative seasonality. Oracle Machine Learning for SQL implements exponential smoothing using a state of the art state space method that incorporates a single source of error (SSOE) assumption which provides theoretical and performance advantages. Access to research results is organized using the APEX web application creation tool. The considered toolkit will help in making decisions when assessing the base of the tax potential of the region and planning tax revenues. The results of the Holt-Winters exponential smoothing model algorithms research are presented and the losses of personal income tax in Ivano-Frankivsk region in 2022 are estimated.