The Economic Value of Deep Learning Forecasts in Electricity Balancing Markets

Description

This package contains the data and code required to reproduce every table and figure in the associated paper, which compares five deep learning architectures and a cointegration-based econometric model on UK half-hourly electricity imbalance settlement prices, and then evaluates each model on the profitability of trading strategies built from its forecasts against a zero-skill benchmark. The dataset covers 1 January 2018 to 31 December 2025 at half-hourly resolution. The imbalance settlement price, wind generation, solar generation, demand and the APX Market Index Price are drawn from the Elexon Balancing Mechanism Reporting Service. The gas price is the National Gas Transmission System Average Price, published through the Office for National Statistics under the Open Government Licence v3.0. Once the rolling and lagged features are constructed the modelling sample runs from 16 January 2018 and contains 139,536 observations.

Steps to reproduce

The package holds twenty Jupyter notebooks numbered in the order they run, the raw source series as downloaded, and the processed dataset the models are trained on. The notebooks build the dataset, run the exploratory and diagnostic analysis, estimate the VECM-GARCH and the five neural architectures under a common harness alongside a gradient boosting baseline, and produce the accuracy and trading evaluations. The README maps every table and figure in the paper to the notebook that generates it. Analysis was carried out in Python 3.13.5 with seeds fixed at 42. Neural network training on GPU is not bitwise deterministic, so reported errors may move in the third decimal.

Institutions

• University College London

  England, London

Categories

Licence