We investigate the performance of various value-at-risk (VaR) models in the context of the highly volatile Nordic power futures market, examining whether simple averages of models provide better results than the individual models themselves. The individual models used are normally distributed GARCH, t-distributed GARCH, t-distributed GJR–GARCH, a quantile regression using RiskMetrics, a quantile regression using t-distributed GARCH, RiskMetrics with Cornish–Fisher and a filtered historical simulation using t-distributed GARCH. We find that RiskMetrics with Cornish–Fisher and normally distributed GARCH perform worse than the other individual models. The average models generally outperform the individual models at a 5% significance level. The conditional independence test reveals that the models are only partially capable of accounting for the volatility clustering of the Nordic power futures. Investors in the Nordic electricity markets should therefore use several methods and average them to be more confident in their VaR estimates.