Large Language Models for Forecasting of Electricity Price Spikes

| Source: arXiv AI

Tags: National Electricity Market, electricity price forecasting, gradient boosting, Brier score, LLM feature extraction

A preregistered-style test on Australia's electricity market found LLM-read market notices did not reliably improve short-horizon price-spike forecasts, a useful negative result for LLM feature extraction.

Details

The authors asked whether market notices read by a local LLM could improve short-horizon price-spike forecasts in five regions of Australia's National Electricity Market. Features were extracted once per notice, and a gradient-boosting model produced onset probabilities. In a July 2022 to July 2024 back-test, notice features did not improve the 60-minute Brier score of a numeric model that already used the operator's next-hour forecasts. At 24 hours they helped, but simple rule-based features did almost as well. On a locally frozen holdout from 15 August to 17 October 2026, none of the hypotheses were established: relative Brier skill was -0.09% at 60 minutes and an inconclusive +0.37% at 24 hours. With only 5 spike onsets on 3 days, the holdout could neither confirm nor rule out a 24-hour effect.