The Methods Landscape for Electricity Price and Spread Forecasting

Electricity price forecasting is the task of predicting the price of electric energy at a given location and time in a wholesale market. Spread forecasting targets the difference between two prices, most commonly the day-ahead versus real-time spread: the gap between the price set in the day-ahead auction and the price realized in real-time operation for the same location and hour. Both problems share a methods toolkit, but they differ sharply in signal-to-noise. Spreads are close to a martingale difference in efficient markets, so most of the achievable value lives in conditioning (knowing when the distribution shifts) and in the tails, not in point accuracy.

This article surveys the model families, feature sets, and evaluation practices that define the current landscape.

Decision target first, model second

A useful forecast is defined by the downstream decision it serves. Candidate targets include hourly day-ahead locational marginal price (LMP, the price at a specific network node that reflects energy, congestion, and losses), five-minute real-time LMP, day-ahead versus real-time spreads, ancillary-service prices (payments for reserve and regulation capability), the probability and magnitude of scarcity or negative prices, quantiles or joint scenario paths, and even the rank of nodes or hours rather than absolute price levels.

The target dictates the loss and the model. A virtual-bidding system, which places financial positions on the spread between day-ahead and real-time prices, may value directional precision and tail calibration far more than root mean squared error. A battery optimizer needs coherent multi-hour price paths, because state of charge couples decisions across time and independent hourly forecasts break that coupling.

Why strong baselines are hard to beat

Electricity price series are heteroskedastic (their variance changes over time), heavy-tailed, strongly seasonal, bounded by market rules such as price caps and floors, and structurally nonstationary as the generation fleet and market design evolve. Under these conditions, simple and robust methods remain difficult to beat consistently:

• seasonal naive and persistence forecasts (repeat the value from the same hour yesterday or last week);
• linear and regularized regression on calendar and fundamental features;
• ARIMA and ARX time-series models, and regime-switching variants that maintain separate dynamics for different market states;
• gradient-boosted trees and random forests;
• two-part designs with separate normal-regime and spike models, where a spike is a short-lived extreme price excursion;
• forecast combinations and residual correction on top of a base model.

Robust transformations and explicit spike treatment are often more valuable than architectural novelty. Any new method should be compared against this baseline set under identical data and evaluation protocols; comparing a heavily tuned deep model against weak baselines is a recognized failure mode.

Deep temporal models: where they pay and where they do not

The deep learning families used in this domain include LSTM and GRU recurrent networks, temporal convolutional networks (TCNs, which model sequences with dilated convolutions), sequence-to-sequence models, temporal fusion transformers, patch-based and long-horizon Transformers, mixture density networks (networks that output distribution parameters), and normalizing flows and diffusion-style generative models.

Transformers can integrate long history and many covariates, which is genuinely useful when the feature space is wide and interactions are complex. But they do not automatically encode dispatch economics or network topology: nothing in the attention mechanism knows about the merit order (the ranking of generators by marginal cost that determines which units set the price) or transmission constraints. Deep models therefore tend to pay when combined with structure, and to disappoint when deployed as opaque sequence models on a nonstationary, spiky series where a tuned tree ensemble already captures most of the learnable signal.

A high-value hybrid pattern decomposes the forecast as a structural component plus a learned residual. The structural component approximates the merit order, unit commitment, or network physics from load, supply, and topology inputs; the residual model captures omitted behavior from history, weather, and regime features. Because the structural part encodes why prices form, this design can outperform an opaque time-series model during distribution shifts.

Probabilistic and distributional models

The production frontier is a calibrated distribution, not a single point estimate. Approaches include quantile regression (directly estimating conditional quantiles), parametric distributions with time-varying parameters, ensembles and bootstrap paths, copulas for cross-node and cross-hour dependence, scenario generators based on stochastic differential equations and jump processes, variational or adversarial generative models, and conformal prediction, a distribution-free wrapper that provides empirical coverage guarantees.

For decisions that couple across time and space, a scenario set must preserve temporal autocorrelation, cross-node correlation, price-load-renewable dependence, joint scarcity regimes, market caps, floors, and negative-price behavior, and the dependence between day-ahead and real-time prices. Independent hourly scenarios that violate path dependence are a common failure mode.

Ensembling

Forecast combination is a baseline technique in its own right, and the recommended frontier is itself an ensemble: a structure-aware combination of fundamental models, topology or graph-neural-network information, probabilistic temporal models, and calibrated joint scenario generation, evaluated by downstream risk-adjusted value rather than accuracy alone.

Feature families

Features group into a few families:

• Fundamentals: load and load forecast errors; fuel and emissions prices; reserve requirements and scarcity indicators.
• Weather and renewables: wind and solar forecasts and their errors; temperature, humidity, wind speed, irradiance; full weather ensembles.
• Topology and outages: generation availability, outages, and derates (temporary capacity reductions); monitored transmission constraints; shift-factor features that describe how injections at nodes affect flows on constrained lines; intertie schedules and neighboring-market prices.
• Market memory and calendar: lagged prices and spreads; calendar, daylight-saving, and holiday effects; rule-regime and market-design version indicators, since market rule changes alter price formation.

Model family comparison

Loss functions

Optimizing mean squared error alone is a mistake for most decisions. Alternatives include MAE or Huber loss for robustness, pinball loss (the asymmetric loss whose minimizer is a conditional quantile) for quantile models, CRPS or the energy score for full distributions, weighted losses that emphasize scarcity intervals, focal or classification losses for spike occurrence, differentiable downstream profit or regret, calibration penalties, and directional or ranking objectives. Decision-focused learning can train forecasts directly against optimization regret, but it must preserve honest out-of-sample evaluation and avoid learning simulator artifacts.

Evaluation and validation

Credible evaluation uses rolling-origin or walk-forward schemes: the model is repeatedly trained on data up to a cutoff and tested on the period after it, never with random train-test splits. Requirements include timestamp-correct feature reconstruction (using only the forecast vintages available at decision time), retraining at a realistic cadence, probabilistic calibration tests, economic backtests that include fees, credit, and fill rules, and stress tests for scarcity events and topology changes.

Regime-sliced testing matters as much as aggregate scores: performance should be reported separately by season, hour, node, and price regime, because a model can look strong on average while failing exactly in the scarcity hours where money is made or lost.

Recognized failure modes include leakage from revised forecasts or finalized outage data, evaluating only system hubs while trading at individual nodes, clipping spikes and then claiming improved accuracy, using symmetric error metrics for asymmetric decisions, and ignoring rule changes and price-cap regimes.

Where the value actually lives

No method predicts prices in any strong sense. Spread forecasting in particular has low signal-to-noise, and improvements in average point accuracy translate weakly into economic value. The durable sources of edge are conditioning (features and regime models that identify when the price distribution has shifted, such as outage, weather, and scarcity indicators) and the tails (calibrated estimates of spike probability and magnitude). This is why the recommended baseline pairs gradient-boosted trees on calendar and fundamental features with a dedicated spike classifier and quantile regression, benchmarked against seasonal-naive and linear models, and why the frontier is judged by downstream risk-adjusted value rather than error metrics.

Glossary

• LMP (locational marginal price): the price of energy at a specific network node, reflecting energy, congestion, and losses.
• Day-ahead / real-time spread: the difference between the day-ahead auction price and the realized real-time price for the same location and hour.
• Virtual bidding: placing purely financial bids that settle on the day-ahead versus real-time spread.
• Merit order: the ranking of generators by marginal cost that determines which unit sets the market price.
• Spike: a short-lived extreme price excursion, often driven by scarcity.
• Derate: a temporary reduction in a generator's available capacity.
• Shift factor: the sensitivity of flow on a transmission constraint to injection at a node.
• Pinball loss: the asymmetric loss function whose minimizer is a conditional quantile.
• CRPS (continuous ranked probability score): a proper scoring rule that evaluates a full predictive distribution against an observed outcome.
• Conformal prediction: a distribution-free method that wraps a model to produce intervals with empirical coverage guarantees.
• Rolling-origin (walk-forward) evaluation: repeatedly training on data up to a cutoff and testing on the subsequent period, mimicking live deployment.
• Calibration: agreement between predicted probabilities or quantiles and observed frequencies.