’žMultikulturowo[. Rynki dnia nastpnego, ekonometria i uczenie maszynowe
[Across cultures. Day ahead markets, econometrics and machine learning]
Carlo Lucheroni, Costantino Ragno
(School of Science and Technology, University of Camerino, WBochy)
As noted in Weron (2014, IJF), most papers on electricity day-ahead prices
come from two mainstream research cultures, econometrics and machine learning
(ML). People from these two groups tend to stick to their own approaches,
perspectives and research networks, finding it difficult to recognize a model
they already use when it is described in the language of the competing field.
In this seminar, a simple model that is well known in econometrics as regime
switching zero-lag autoregression is shown to be well known in the ML community
as well, under the name of hidden Markov mixture model. Using results from ML
research, it will be shown how such a simple model can do unsupervised (deep)
learning and data clustering, organizing and parametrizing market data in a
very clear and efficient way. Once estimated, this model is able to synthetically
generate time series with features like concurrent night/day seasonality and
spiking, and spike clustering, impressively similar to original data. It will be
also shown that this model lends itself to analytic probabilistic forecasting
in a natural way.