Μοντέλα πιθανοκρατικής πρόβλεψης ενδο-ημερήσιων τιμών φυσικού αερίου στον κόμβο TTF και αξιολόγηση απόδοσης μέσω PnL

Βενιζέλος, Νικόλαος

Master Thesis

Author

Βενιζέλος, Νικόλαος

Date

2026-05

Abstract

This thesis investigates the fundamental question of whether modern machine learning and deep learning models can outperform the simple persistence baseline in forecasting TTF natural gas prices at ultra-short horizons (10-60 minutes), and whether robabilistic forecasts offer practical value for risk management and trading applications. The study also examines whether the geopolitical instability of 2025 created exploitable predictive patterns in the market. A total of 147,983 intraday minute-level prices from 2025 were analyzed using 8 forecasting models: Naive Persistence, Ridge Regression, Random Forest, Quantile Gradient Boosting, SVR, XGBoost, LSTM, and QRNN. Feature engineering produced 77 input variables including lagged values, rolling statistics, and technical indicators. Evaluation combined a static train/validation/test split with rolling window retraining, assessed through probabilistic metrics (CRPS, prediction intervals) and trading simulation. The key findings strongly validate the Efficient Market Hypothesis. The simple persistence model outperforms all complex models (MAE €0.0554 versus €0.0571–€2.98), with a firstorder autocorrelation coefficient ρ₁=0.9996 indicating random walk behavior. Despite being unable to beat the baseline in point forecasts, probabilistic models proved valuable for uncertainty quantification, with SVR achieving well-calibrated prediction intervals (92.94% coverage) and a CRPS score of €0.1629. Rolling window retraining dramatically improved treebased model performance (72–73%) but still failed to surpass the baseline. All trading strategies proved unprofitable after transaction costs (negative Sharpe ratios, win rates ∼40%), confirming that low forecasting error does not translate to profitability in efficient markets. The thesis contributes scientific validation of the Efficient Market Hypothesis on ultra-highfrequency data during extreme events, a comprehensive evaluation framework combining probabilistic metrics with financial simulation, and evidence that "negative results" carry theoretical value. Practical applications include the design of realistic forecasting systems for liquid energy markets, the use of probabilistic forecasts for risk management rather than directional position-taking, and reference benchmarks for future research. The central conclusion emphasizes: "the market is smarter than our models – and that is precisely what theory predicts".

Postgraduate Studies Programme

Πληροφοριακά Συστήματα και Υπηρεσίες

Department

Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών Συστημάτων

Number of pages

Language

Greek

URI

https://dione.lib.unipi.gr/xmlui/handle/unipi/19635

Collections

Τμήμα Ψηφιακών Συστημάτων

Show full item record

Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα