A Machine Learning approach to Risk Minimisation in Electricity Markets with Coregionalized Sparse Gaussian Processes

Daniel Poh, Stephen Roberts, Martin TegnérarXiv 2019 · DOI · arXiv

Abstract

The non-storability of electricity makes it unique among commodity assets, and it is an important driver of its price behaviour in secondary financial markets. The instantaneous and continuous matching of power supply with demand is a key factor explaining its volatility. During periods of high demand, costlier generation capabilities are utilised since electricity cannot be stored and this has the impact of driving prices up very quickly. Furthermore, the non-storability also complicates physical hedging. Owing to these, the problem of joint price-quantity risk in electricity markets is a commonly studied theme. We propose using Gaussian Processes (GPs) to tackle this problem since GPs provide a versatile and elegant non-parametric approach for regression and time-series modelling. However, GPs scale poorly with the amount of training data due to a cubic complexity. These considerations suggest that knowledge transfer between price and load is vital for effective hedging, and that a computationally efficient method is required. To this end, we use the coregionalized (or multi-task) sparse GPs which addresses the aforementioned issues. To gauge the performance of our model, we use an average-load strategy as comparator. The latter is a robust approach commonly used by industry. If the spot and load are uncorrelated and Gaussian, then hedging with the expected load will result in the minimum variance position. Our main contributions are twofold. Firstly, in developing a coregionalized sparse GP-based approach for hedging. Secondly, in demonstrating that our model-based strategy outperforms the comparator, and can thus be employed for effective hedging in electricity markets.

Introduction

Electricity is a unique commodity. Unlike oil, wheat, or gold, it cannot be stored at scale, requiring an instantaneous and continuous balance between supply and demand. This physical constraint leads to extreme price volatility: when demand spikes, utilities must activate expensive "peaker" plants, causing prices to skyrocket in seconds. For electricity retailers, this creates a significant joint price-quantity risk. They are most vulnerable when prices are high and their customers’ consumption is at its peak.

Traditional risk management in this sector often relies on "average-load" strategies. These benchmarks typically assume that price and load are either uncorrelated or follow simple Gaussian distributions. In reality, the relationship between these variables is non-linear, seasonal, and highly volatile. This paper proposes a machine learning framework using Coregionalized Sparse Gaussian Processes (CSGPs) to model these dependencies and optimize hedging strategies.

Methodology

To address the complexities of electricity markets, we employ two sophisticated extensions of standard Gaussian Processes (GPs):

1. Coregionalization (Multi-task Learning)

Instead of modeling price and load as independent variables, a coregionalized GP treats them as related tasks. This multi-task approach allows the model to "transfer knowledge" between variables, capturing the underlying drivers—such as weather patterns or time-of-day cycles—that influence both price and load simultaneously.

2. Sparse Approximations (DTC)

Standard GPs suffer from $O(n^3)$ computational complexity, making them impractical for the high-frequency datasets common in energy markets. We utilize the Deterministic Training Conditional (DTC) approach. By using a small set of "inducing points" to approximate the full dataset, we drastically reduce the computational burden without sacrificing the model's ability to capture essential trends.

The Composite Covariance Kernel

To capture the idiosyncratic behavior of electricity data, we constructed a composite kernel designed to handle multiple temporal scales:

  • Squared Exponential: Captures broad, smooth trends.
  • Matérn 5/2: Accommodates non-smooth, "jagged" data points.
  • Periodic (12h, 24h, 168h): Accounts for daily and weekly seasonality.
  • Rational Quadratic: Models medium-term irregularities and multi-scale variations.

Empirical Evaluation

We tested our model using UK electricity market data from 2016 to 2018. This included hourly day-ahead spot prices from Nord Pool and forward contract prices from Bloomberg. Since public hourly consumption data was unavailable, load was estimated using 30-minute power demand snapshots from the National Grid ESO.

Key Insights

  • Sparsity as a Low-Pass Filter: Interestingly, high sparsity (using only 1% of data as inducing points) often improved hedging performance. This suggests that sparsity acts as a natural filter, preventing the model from overfitting to short-term noise and price spikes, and forcing it to focus on the stable, long-term trends relevant for monthly hedging.
  • The Failure of the "Oracle" Benchmark: We compared our model against an "average-load" strategy granted "oracle" knowledge (the actual future average load). Even with this advantage, the industry-standard approach failed because it could not account for the critical covariance between price and load.
  • The Importance of Recency: Models trained on a single month of data (CSGP-1M) outperformed those with longer training windows (two or three months). In fast-moving electricity markets, older data can introduce "stale" correlations that degrade the accuracy of current forecasts.

Results

The performance of the CSGP-1M model over the three-year test period was substantial. While the standard industry benchmark (Average Load) resulted in a total loss of £3.61 million, our CSGP-based strategy delivered a cumulative payoff of £5.42 million. The model demonstrated remarkable consistency, outperforming the benchmark in the majority of months across varying market conditions.

Conclusion and Future Work

This research demonstrates that Coregionalized Sparse Gaussian Processes provide a robust framework for managing joint price-quantity risk in electricity markets. By capturing the complex dependencies between price and load, the CSGP approach offers a significant advantage over traditional industry benchmarks.

Future research will explore dynamic hedging, where positions are adjusted daily as new data arrives, rather than the static monthly hedges used in this study. Additionally, we aim to investigate fully Bayesian treatments of kernel parameters to better quantify uncertainty and improve the model's handling of extreme "tail risk" events.

Summary updated

Abstract

The non-storability of electricity makes it unique among commodity assets, and it is an important driver of its price behaviour in secondary financial markets. The instantaneous and continuous matching of power supply with demand is a key factor explaining its volatility. During periods of high demand, costlier generation capabilities are utilised since electricity cannot be stored and this has the impact of driving prices up very quickly. Furthermore, the non-storability also complicates physical hedging. Owing to these, the problem of joint price-quantity risk in electricity markets is a commonly studied theme. We propose using Gaussian Processes (GPs) to tackle this problem since GPs provide a versatile and elegant non-parametric approach for regression and time-series modelling. However, GPs scale poorly with the amount of training data due to a cubic complexity. These considerations suggest that knowledge transfer between price and load is vital for effective hedging, and that a computationally efficient method is required. To this end, we use the coregionalized (or multi-task) sparse GPs which addresses the aforementioned issues. To gauge the performance of our model, we use an average-load strategy as comparator. The latter is a robust approach commonly used by industry. If the spot and load are uncorrelated and Gaussian, then hedging with the expected load will result in the minimum variance position. Our main contributions are twofold. Firstly, in developing a coregionalized sparse GP-based approach for hedging. Secondly, in demonstrating that our model-based strategy outperforms the comparator, and can thus be employed for effective hedging in electricity markets.
BibTeX
@misc{poh2019machine,
  author = {Daniel Poh and Stephen Roberts and Martin Tegnér},
  title = {{A Machine Learning approach to Risk Minimisation in Electricity Markets with Coregionalized Sparse Gaussian Processes}},
  year = {2019},
  eprint = {1903.09536},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1903.09536},
}
Figure