Seminarium QFRG i DSLab „Forecasting of the day-ahead energy prices” [17.01.2022]
Serdecznie zapraszamy do wzięcia udziału w kolejnym seminarium z cyklu spotkań organizowanych przez QFRG (Quantitative Finance Research Group) oraz DSLab (Data Science Lab). Podczas spotkania Maciej Przybyła, absolwent podyplomowych studiów WNE UW „Data Science w zastosowaniach biznesowych. Warsztaty z wykorzystaniem programu R” oraz dr hab. Piotr Wójcik, prof. ucz. z Katedry Finansów Ilościowych WNE UW przedstawią badanie „Forecasting of the day-ahead energy prices”.
Badacze w nowatorski sposób prognozują ceny energii, uwzględniając udział elektrowni wiatrowych i słonecznych w jej produkcji, emisję CO2 oraz ceny węgla. W badaniu zastosowano model statystyczny Lasso Estimated AutoRegressive (LEAR), model uczenia maszynowego SVM (Support vector machine) oraz ich kombinację .
Seminarium odbędzie się 17 stycznia 2022 r. od godz. 15:00 za pośrednictwem platformy Google Meet.
Link do spotkania: https://meet.google.com/dfx-mfwc-whg
--
Prices of energy have impact on all individuals – households and entrepreneurs, as they influence prices of all other goods. However, energy prices forecasting (EPF) lacks a rigorous approach to compare and to evaluate new forecasting models. Lago et al. (2021) analyzed different factors affecting the quality of the research and proposed a set of guidelines, which we follow. Most EPF models concentrate on historical prices of energy, sometimes using also the data on generation (supply) and load (demand) of energy as additional predictors. The main novelty of our approach is the incorporation on additional predictors influencing energy prices (share of wind and solar plants in power generation, CO2 emissions and coal prices). The main objective of our research is the prediction of the day-ahead energy prices in the Polish market. We apply a statistical model called Lasso Estimated AutoRegressive (LEAR), a selected machine learning model (SVM) and their combination (hybrid). In addition, we compare the performance of Facebook’s Prophet model and a time series regression with no variable selection. We find that including a wider set of theoretically justified factors allows to obtain more accurate predictions of the price of energy. Regular model recalibration gives more accurate predictions and 1-, 2-weekly recalibration gives similar forecast accuracy as daily adjustments. As expected, longer training sample gives significantly more accurate predictions.
