Probabilistic forecasting in energy: how to read price risk

Consider two deliberately simplified forecasts for a single delivery hour. Forecast A assigns a 50% chance each to $40/MWh and $60/MWh. Forecast B assigns a 90% chance to $20/MWh and a 10% chance to $320/MWh. Both means are $50/MWh. These are teaching examples, not market observations or model outputs.

For a buyer with 100 MWh of unhedged consumption in that hour, both forecasts imply an expected energy cost of $5,000. But their possible bills are very different. Under A, the bill is either $4,000 or $6,000. Under B, it is either $2,000 or $32,000.

Suppose that buyer could fix the price at $60/MWh, with no additional fees in this example. Fixing costs $6,000: $1,000 above the expected spot bill in either case. Whether that premium is worthwhile depends on the buyer’s ability to absorb the high-cost outcome. A buyer concerned only with expected cost could make the same choice in both cases. A buyer with a binding cash limit might not.

A probability distribution does not make the decision for you. It reveals the tradeoff that a single number conceals.

Electricity exposure is specific to a location and delivery interval. A locational marginal price reflects the incremental cost of serving demand at a particular point on the grid, and its components can include congestion and losses. A forecast for a regional hub therefore does not automatically describe the price at an individual asset. EIA definition of locational marginal price

Prices can also be negative. The US Department of Energy has documented how wind generation affects the timing and location of prices, including negative-price periods near renewable plants. A useful price distribution must accommodate the outcomes relevant to its market rather than assume a positive, symmetric range around an average. DOE analysis

Specify the product before interpreting the uncertainty: day-ahead or real-time, hub or node, five-minute interval or hourly average, and the forecast's lead time. Those choices define what is being predicted and what observation will eventually count as the outcome.

The same reasoning extends beyond electricity prices. Demand, wind output, solar generation, and the timing of their changes all affect decisions before their outcomes are known. Probabilistic energy forecasting spans these different targets; the Global Energy Forecasting Competition 2014 included separate load, price, wind, and solar tracks. Research overview

The useful question depends on the work. An electricity buyer may care about a price threshold. A renewable generator may care about the size of a production shortfall. A battery operator needs to consider the sequence of prices alongside stored energy, efficiency, power limits, and degradation.

That last case contains an important distinction. A probability distribution for each hour does not, by itself, describe how outcomes move together across hours. A high-price event lasting fifteen minutes and one lasting four hours have different implications for storage. Decisions across time require information about sequences and dependence as well as individual intervals.

A point forecast compresses the outlook into one statistic. That might be the mean, or it might be the median; the label matters. In Forecast B above, the expected price is $50/MWh, while the median is $20/MWh.

Quantiles describe positions within the forecast distribution. P50 is the median. P10 and P90 mark the lower and upper boundaries of a central 80% prediction interval. For a continuous distribution, the model assigns about 10% probability below P10 and 10% above P90. Those are probability statements about the forecast, not guarantees about the next observation. Prediction intervals explained

A prediction interval describes a future observation. A confidence interval typically describes uncertainty in an estimated parameter, such as a mean. They are different objects. Also check the convention behind a percentile label: this article uses P90 to mean the 90th percentile of the forecast distribution, or a value with about 90% probability below it for a continuous distribution.

A fan chart draws nested intervals over time. A density view shows where probability concentrates, making it easier to see an asymmetric outlook or several distinct price ranges. A narrow, bright region and a broad upper tail communicate different things even when the median barely moves.

Thresholds turn that picture into a practical question: what probability does the forecast assign to a price above $100/MWh, or below zero? The threshold should come from the decision being considered. In our discrete example, the probability of exceeding $100/MWh is zero under A and 10% under B.

Density needs a little care: the height or brightness at a single price is not itself the probability of that exact price. Probability belongs to a range, represented by area under a continuous density or by probability mass in discrete buckets. Likewise, an 80% interval at each future hour does not imply an 80% chance that the entire day's path stays inside all the intervals.

Several method families can produce probabilistic forecasts. Quantile regression estimates specified percentiles directly. Distributional models estimate a predictive distribution. Ensembles generate multiple possible outcomes, which can then be calibrated against observations. These are choices about how to represent uncertainty; they do not all require a large neural network. Review of probabilistic electricity price forecasting

Modern AI expands the ways we can learn forecast distributions from data. Neural models can learn relationships across many time series, and pretrained forecasting models can transfer some of that learning to new series. Chronos, for example, demonstrated an approach that adapts transformer architectures to time-series forecasting and generates probabilistic forecasts by sampling future trajectories. Chronos research

The opportunity in energy is to make a distribution sensitive to the information available when the forecast is issued. The outlook should be able to change in shape as well as level.

But “AI” does not establish that the probabilities are dependable. That has to be measured against subsequent outcomes. The useful distinction is between a model that produces probabilities and a model whose probabilities have been shown to work for the market, horizon, and decision in question.

Start with calibration. Across enough comparable forecasts, does an 80% interval contain the outcome about 80% of the time? Check this by horizon and relevant operating conditions as well as in aggregate. A reassuring overall result can conceal a weakness in the hours that matter most. Calibration and sharpness

For example, if actual prices fall inside a model's P10–P90 range in only 620 of 1,000 matched intervals, observed coverage is 62%, compared with the nominal 80%. That is a useful warning about undercoverage in that sample. It is not a complete diagnosis: outcomes across neighbouring intervals can be correlated, and conditions may have changed.

Then ask how informative the forecast is. An extremely wide interval can capture many outcomes while offering little help. Forecast evaluation should reward useful concentration and penalize misses appropriately. Quantile loss evaluates specified quantiles; interval scores account for both interval width and outcomes outside the interval. Evaluating distributional forecasts

Compare distribution scores on the same targets and units against a relevant baseline. Report tail-event results separately when rare spikes are central to the use case; a good average score can still hide poor performance where exposure is largest. Probabilistic forecasting methods and evaluation

The record must also preserve what was known at issuance. Comparing an earlier outcome with a forecast updated after that outcome became known would not measure the decision the original user faced. Issue time, target interval, model version, matched observations, and missing data belong beside the result.

Finally, evaluate the decision itself. A statistical improvement and an operational improvement are separate claims. A dispatch or purchasing strategy needs its own evaluation with the costs, constraints, and information that would actually have applied.

Our probabilistic forecasting model, Metis 1 Preview, brings this way of reading prices into Ask the Grid. At supported pricing locations, the forward outlook offers Shape and Fan views: density shading for the distribution and nested ranges for uncertainty over time.

In Shape view, look at where the probability concentrates and how the distribution spreads upward or downward. In Fan view, follow the median and the nested prediction intervals. Cleared day-ahead prices are a separate market series; they are not the model’s prediction or an observation of real-time prices.

Open the series details to inspect issuance, model build, units, and horizon. Then use settlement review to compare completed intervals with earlier forecast vintages. Read the matched sample size before drawing a conclusion: a short or incomplete window cannot establish reliable calibration.

This is how we think AI forecasts should enter an energy workflow: as an inspectable distribution connected to its source, timing, and eventual outcome. The model supplies the outlook; the reader can examine what it means for the decision at hand.

Before using a forecast operationally, ask its provider for a matched evaluation at your location and horizon, historical issuances available before delivery, distribution scores against a baseline, and a clear account of missing intervals. For storage or portfolio optimization, also ask whether the output includes joint scenarios or only per-interval quantiles.

A useful forecast helps you prepare for more than the central case. It shows which outcomes deserve attention, how much probability they carry, and what changes when the next piece of information arrives.

Explore Metis 1 Preview or read the forecasting guide.