Why Power Prices Are Harder to Forecast, and Why Fundamental

Forecasting electricity prices has always been difficult. In 2026 it has become harder still. Power markets are being pulled in opposite directions at once: more hours of very low or negative prices as renewables grow, and sharp spikes whenever gas sets the price during periods of stress. For traders, asset owners and investors, the cost of a poor forecast has risen accordingly.

Why prices are harder to predict

Renewables are reshaping the daily price curve

Rapid growth in solar and wind has deepened the so-called duck curve: more low and negative prices around midday and steeper evening peaks. According to analysis by consultancy Ricardo, EU day-ahead markets recorded more than 1,200 negative-price hours in the first quarter of 2026, more than double the same period a year earlier, led by Spain, Portugal and Greece. In Germany, around one in six hours in April 2026 cleared below zero, and those hours coincided with the largest gaps between forecast and actual renewable output.

Gas still sets the price when it matters

Even as gas sets the price for fewer hours, it remains decisive at times of system stress. The European Environment Agency found that in early 2026 gas set the electricity price in around two-thirds of analysed hours in Italy and Poland, but in fewer than one in ten hours in Spain, producing very different price levels. The disruption to LNG flows through the Strait of Hormuz has pushed European gas prices to their highest level since late 2022, feeding directly into power prices in gas-dependent markets.

Structural change breaks historical patterns

New interconnectors, battery storage, flexible demand, market reforms and policy changes all alter how prices form. Patterns that held for years can disappear within months, which undermines any approach that relies purely on history.

The two main forecasting approaches

Statistical and machine-learning models

Statistical models, from regression and time-series methods to modern machine learning, learn relationships from historical data such as recent prices, load, wind and solar forecasts and fuel prices. They are fast and often very accurate for short horizons such as day-ahead and intraday, because they capture autocorrelation, seasonality and recent market behaviour well.

Their weakness is regime change. Academic research on European markets has shown that models relying on lagged prices tend to carry forward outdated conditions, for example gas price levels that no longer apply, and can perform poorly around sudden spikes and drops. They also struggle to explain why a forecast looks the way it does, and cannot easily answer what-if questions about new plants, policies or fuel prices.

Fundamental models

Fundamental models, built in tools such as PLEXOS, BID3 and PROMOD, simulate the power system itself: demand, the merit order of generators, fuel and carbon costs, renewable output, storage, interconnection and network constraints. Prices emerge from how the system must be dispatched to meet demand.

This brings clarity that statistical models cannot. A fundamental model explains which plant is setting the price and why. It can test scenarios such as a prolonged gas shock, a faster solar build-out, a new interconnector or a change in market rules. It is therefore essential for medium- and long-term forecasting, asset valuation and investment decisions, where history is a poor guide to the future.

Its limitations are effort and short-term precision: fundamental models need careful data and calibration, and they may not capture behavioural effects, bidding strategies or very short-term dynamics as precisely as a well-tuned statistical model.

Why combining them gives the best answer

The strengths of the two approaches are complementary, and leading practitioners increasingly use them together:

Fundamental structure, statistical precision: the fundamental model sets the price level and structure based on system conditions, while statistical models refine the short-term shape using recent behaviour.
Fundamental outputs as statistical inputs: modelled marginal costs, residual load and scarcity indicators make powerful features for machine-learning models, helping them recognise regime changes quickly instead of relearning from lagged prices.
Statistical correction of fundamental bias: statistical models can learn and correct systematic differences between fundamental results and observed prices, such as bidding mark-ups or scarcity pricing.
Ensembles across horizons: statistical models dominate the first hours and days, with weight shifting progressively to the fundamental model for weeks, months and years ahead.
Probabilistic forecasts: combining scenario ranges from the fundamental model with statistical uncertainty estimates gives a realistic distribution of outcomes, which is what risk management and asset valuation actually need.

What good looks like in practice

Start from the decision the forecast supports, whether trading, hedging, valuation or planning, and choose horizons accordingly.
Build a well-calibrated fundamental model with transparent, documented assumptions.
Develop statistical models for short-term horizons, using fundamental outputs as features.
Blend and back-test systematically, monitoring performance by hour, season and market regime.
Automate data and runs, and deliver forecasts into trading and risk systems so they are actually used.

Orivyn's optimisation and forecasting practice helps clients build exactly this kind of capability, from market modelling and price forecasting to model implementation and support, and integration with ETRM and CTRM platforms.

Sources

Market figures are as reported by the sources above at the time of writing and will change; they are provided for context, not as investment advice.

Why Power Prices Are Harder to Forecast, and Why Fundamental and Statistical Models Work Best Together