Using neural networks and extreme value distributions to model electricity pool prices: Evidence from the Australian National Electricity Market 1998–2013

• Neural nets are unable to properly capture spiky price behavior found in the electricity market.
• We modeled electricity price data from the Australian National Electricity Market over 15 years.
• Neural nets need to be augmented with other modeling techniques to capture price spikes.
• We fit a Generalized Pareto Distribution to price spikes using a peaks-over-thresholds approach.

Competitors in the electricity supply industry desire accurate predictions of electricity spot prices to hedge against financial risks. Neural networks are commonly used for forecasting such prices, but certain features of spot price series, such as extreme price spikes, present critical challenges for such modeling. We investigate the predictive capacity of neural networks for electricity spot prices using Australian National Electricity Market data. Following neural net modeling of the data, we explore extreme price spikes through extreme value modeling, fitting a Generalized Pareto Distribution to price peaks over an estimated threshold. While neural nets capture the smoother aspects of spot price data, they are unable to capture local, volatile features that characterize electricity spot price data. Price spikes can be modeled successfully through extreme value modeling.