Neural Networks and Deep Learning

Lecture 13 · Game: Higher or lower than today?

Higher or lower than today?

You are the trader of a Romanian electricity supplier. You see the daily average day-ahead price of the last three weeks on the Romanian market (OPCOM, Romanian electricity and gas market operator) and decide: will tomorrow's average price be higher or lower than today's? After each decision you see the forecast of the foundation model of Lecture 13 and the real price. All 10 days come from the test year of the lecture (1 November 2025 to 31 October 2026).

The model is Chronos-2, a time-series foundation model pre-trained by Amazon on many other series and used zero-shot: it was never trained on Romanian prices. It reads the last 2,048 hourly prices (about 85 days) and forecasts the 24 hours of tomorrow; the game averages them. For comparison you also see LEAR (an autoregressive model estimated with the LASSO, least absolute shrinkage and selection operator), the classical benchmark of electricity price forecasting, recalibrated every day on three years of data.

Class board for the projector

Data: Romanian day-ahead prices (OPCOM, via energy-charts.info), daily averages of the 24 hourly prices; forecasts: cached test-year forecasts of analysis_L13.py (lectures/L13/results/forecasts_hourly.csv). The 40 days of the game are drawn at random from the 360 evaluation days (seed 20261108); every game draws 10 of them. Code: lectures/L13/make_game_L13.py.