Leapfrog

Price Modelling Fundamentals

REPORT ID: 03-PRICE-MODELLING · 5 MIN READ

What drives electricity prices

Electricity prices are determined by the interaction of supply and demand under physical constraints. The main drivers can be grouped into several categories:

Demand-side drivers

  • Temperature: Heating demand in winter and cooling demand in summer create strong seasonal and intra-day load patterns.
  • Economic activity: Industrial production and commercial activity influence total consumption, though the effect is gradual.
  • Calendar effects: Weekdays vs. weekends, public holidays, and seasonal patterns create predictable demand shapes.
  • Time of day: Consumption follows a daily profile with a morning ramp, midday plateau, and evening peak.

Supply-side drivers

  • Fuel prices: Natural gas, coal, and oil prices directly affect the marginal cost of thermal generation through the merit order.
  • Carbon price: The EU Emissions Trading System (ETS) carbon price adds a cost to CO₂-emitting generation, raising the marginal cost of coal and gas plants.
  • Renewable generation: Wind speed and solar irradiance forecasts determine the volume of near-zero marginal cost generation available.
  • Plant availability: Scheduled maintenance, forced outages, and nuclear availability affect the supply stack.
  • Hydro reservoir levels: In markets with significant hydropower (e.g., Nordics, Alps), reservoir fill levels strongly influence supply and prices.

Network and market structure

  • Cross-border flows: Interconnector capacities and price differentials between zones drive imports and exports.
  • Grid congestion: Transmission constraints can create price separation between zones even within a single country.

Modelling approaches

Fundamental models

Fundamental (or structural) models simulate the physical dispatch process. They construct a merit order from known generation capacities, fuel costs, and carbon prices, then clear supply against a demand forecast.

Strengths: Interpretable, can incorporate specific plant-level information, handle scenario analysis well (e.g., nuclear outages, fuel price shocks).

Limitations: Require detailed input data (plant parameters, fuel curves), computationally expensive for large portfolios, and sensitive to input assumptions.

Statistical and time-series models

Statistical models treat price as a time series and use historical patterns to forecast future values. Common approaches include:

  • ARIMA / SARIMA: Autoregressive models capturing trend, seasonality, and autocorrelation.
  • Regression models: Use explanatory variables (demand forecast, wind forecast, gas price) as features. Often combined with autoregressive error terms (ARX models).
  • Exponential smoothing (ETS): State-space models for trend and seasonal decomposition.

Strengths: Relatively simple to implement, fast to train, and effective for short-term horizons where historical patterns are stable.

Limitations: Struggle with structural breaks, regime changes (e.g., new market rules), and extreme events.

Machine learning approaches

Machine learning methods have become increasingly popular for electricity price forecasting:

  • Gradient-boosted trees (XGBoost, LightGBM): Effective at capturing non-linear relationships between features and prices. Widely used in day-ahead price forecasting competitions.
  • Neural networks: LSTM and transformer architectures can model complex temporal dependencies, though they require more data and tuning.
  • Ensemble methods: Combining several model types often outperforms any single model, especially when individual models capture different aspects of price behaviour.

Strengths: Can capture complex non-linear interactions, handle many features, and often achieve strong out-of-sample performance.

Limitations: Less interpretable, risk of overfitting (especially with small datasets), and require careful validation.

Hybrid approaches

In practice, many forecasting systems combine elements from multiple approaches:

  • A fundamental model provides a base price scenario.
  • A statistical or ML model corrects the fundamental model’s residuals.
  • Expert overrides handle known events (planned maintenance, extreme weather).

This layered approach leverages the interpretability of fundamental models and the pattern-recognition capability of data-driven methods.

Building a forecasting pipeline

Step 1: Define the target

Be specific about what you are forecasting. Day-ahead clearing prices for a specific bidding zone (e.g., DE-LU) for each hour of the next day is a common target. Intraday VWAP or imbalance prices require different approaches and data.

Step 2: Collect input data

Key data sources include:

  • Historical prices: ENTSO-E Transparency Platform, power exchange APIs (EPEX SPOT, Nord Pool).
  • Demand forecasts: TSO load forecasts, ENTSO-E.
  • Renewable forecasts: Wind and solar generation forecasts from TSOs or weather model providers.
  • Fuel prices: ICE, EEX, and TTF hub prices for gas; API2 for coal; EUA futures for carbon.
  • Calendar data: Day of week, hour, holiday flags, daylight saving transitions.

Step 3: Feature engineering

Transform raw inputs into features the model can use effectively:

  • Lagged prices (price at the same hour yesterday, same hour last week).
  • Rolling averages of demand and renewable generation.
  • Interaction terms (e.g., demand × gas price).
  • Cyclical encoding of time features (hour, month) using sine/cosine transforms.

Step 4: Train and validate

  • Train/test split: Respect temporal ordering. Never use future data to predict past data.
  • Walk-forward validation: Retrain the model periodically on expanding or rolling windows and evaluate out-of-sample performance.
  • Metrics: Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are standard. Also consider metrics that capture directional accuracy and spike detection.

Step 5: Monitor in production

  • Track forecast accuracy over time. Performance degradation may signal data quality issues or structural market changes.
  • Set up alerts for extreme residuals or missing input data.
  • Retrain models regularly with fresh data.

Common pitfalls

  • Lookahead bias: Accidentally including information from the future in training features. Always verify data timestamps carefully.
  • Overfitting to recent history: Markets change. A model trained only on calm periods will fail during price spikes or structural transitions.
  • Ignoring market structure: Treating electricity like a financial time series without accounting for physical constraints (ramping, storage, interconnectors) limits model accuracy.
  • Neglecting negative prices: Increasing renewable penetration means negative prices occur more frequently. Models must handle these correctly.

Summary

Electricity price modelling draws on an understanding of physical market mechanisms and quantitative forecasting techniques. The most effective approaches combine domain knowledge (merit order, fuel costs, renewable intermittency) with data-driven methods capable of capturing complex patterns. A disciplined pipeline — from feature engineering through walk-forward validation to production monitoring — is essential for building forecasts that are reliable in practice.