Comparing GARCH and NN for Forecasting TTF spot Volatility
Translated title
Sammenligning af GARCH og NN modeller til forudsigelse af TTF spot volatilitet
Author
Sørensen, Nicolaj Høgh
Term
4. semester
Publication year
2023
Submitted on
2023-05-30
Abstract
This thesis evaluates whether traditional econometric or machine learning models better predict volatility in TTF gas spot prices. Using historical spot data, a naive random walk benchmark, GARCH and GJR-GARCH are compared with neural networks (LSTM and bidirectional LSTM) across 5-, 10-, and 22-day interval windows. The study employs univariate model setups, computes realized volatility, and assesses predictive performance with RMSPE, RMSE, and the Diebold–Mariano test to evaluate significant forecast differences. Standard pre-estimation checks for GARCH (stationarity, autocorrelation, ARCH effects, and normality) are conducted. Findings indicate that while GARCH-based models improve modestly with longer windows, they generally do not outperform alternatives, whereas LSTM models—particularly the univariate LSTM—consistently deliver the most accurate and robust forecasts across all windows, outperforming both GARCH and the naive benchmark. The results highlight machine learning’s advantage in capturing complex, nonlinear patterns in gas volatility and recommend further refinement of ML approaches to support better decision-making in gas markets.
Denne afhandling undersøger, hvorvidt traditionelle økonometriske modeller eller maskinlæringsmodeller bedst forudsiger volatilitet i TTF gas spotpriser. Med udgangspunkt i historiske spotdata sammenlignes en naiv random walk-benchmark, GARCH og GJR-GARCH med neurale netværk (LSTM og bidirektionel LSTM) på tværs af 5-, 10- og 22-dages intervalvinduer. Arbejdet anvender univariate modelopsætninger, beregner realiseret volatilitet og evaluerer forudsigelserne ved hjælp af RMSPE, RMSE og Diebold–Mariano-test for at vurdere betydningsfulde forskelle. Forud for GARCH-estimation gennemføres standard prætests for stationaritet, autokorrelation, ARCH-effekter og normalitet. Resultaterne viser, at GARCH-baserede modeller får en moderat forbedring ved længere intervalvinduer, men generelt ikke overgår de øvrige modeller, mens LSTM og især den univariate LSTM konsekvent leverer de mest præcise og robuste prognoser på tværs af alle vinduer og outperformer både GARCH og den naive benchmark. Afhandlingen peger dermed på maskinlæringens styrke til at fange komplekse og ikke-lineære mønstre i gasvolatilitet og anbefaler yderligere udvikling og finjustering af ML-tilgange for at forbedre beslutningstagning i gasmarkeder.
[This apstract has been generated with the help of AI directly from the project full text]