Forecasting electricity prices has become increasingly difficult in today’s fast-changing energy markets. Prices can shift quickly due to renewable energy production, demand fluctuations, and regulatory changes. Traditional forecasting models, which rely on years of historical data, often struggle to keep up with these rapid changes.
A new study, titled Less Is More: Data-Driven Day-Ahead Electricity Price Forecasting with Short Training Windows, introduces a different approach: instead of using long-term historical data, it focuses on short training windows of just 7 to 90 days. This allows models to respond more effectively to the most recent market conditions.
The research, conducted by the National Technical University of Athens (NTUA) within the CRETE VALLET framework, compares several machine learning models, including deep learning (LSTM) and widely used boosting models such as XGBoost, LightGBM, and CatBoost, across three European electricity markets which represents different energy systems: Greece, Belgium, and Ireland.
The models were trained using real-world forecast data from ENTSO-E, ensuring that results reflect actual operational conditions faced by energy market participants.
The results are clear: LightGBM consistently delivers the best performance. It achieves higher accuracy and reliability than both deep learning and other machine learning models. The study also finds that training windows of 45 to 60 days provide the best balance, as this timeframe is long enough to capture useful patterns but short enough to avoid outdated or irrelevant data.
In addition, LightGBM shows a strong ability to detect critical factors for traders, system operators, and energy planners such as seasonal trends, price spikes, and extreme market events.
In modern electricity markets, even small improvements in forecasting can lead to better financial decisions and reduced risk. This research shows that simpler, faster models, when trained on recent data, can outperform more complex approaches.
—
Picture by David Levêque on Unsplash
