The impact of the size of the training set on the predictive abilities of neural models on the example of the Day-Ahead Market System of TGE S.A.

Abstract

The main object of the research was to examine the acceptable time horizon that could be predicted by previously learned models of the Day-Ahead Market (DAM) TGE S.A. system. The article contains the results of research on the predicting ability of different ANN models of the DAM TGE S.A. The research was conducted based on data covering the operation of the Polish stock exchange in the period from 2002 to 2019 (the first half of the year). The research was carried out based on the learned ANN models of the DAM system. Data were taken for examination covering the time from 2002 to 2019 (1st half of the year) and was divided into a different period, i.e., a month, a quarter, and a half-year. , year, etc. The MSE, MAE, MAPE, and R2 were adopted as the criteria for assessing the ability of individual models to predict electricity prices. The research was carried out by successively expanding forecasting periods in a rolling manner. For example, for a half-year, prediction time intervals were increased from one week to month, two months, quarter, half-year, etc. results for a model representing a given period. A lot of interesting research results were obtained.