Understanding Ensemble Spread and P10-P90

An ensemble forecast contains multiple members representing different possible weather outcomes.

Each member begins from slightly different initial conditions or model assumptions. These small differences develop over time, producing a range of possible forecasts.

The ensemble helps answer two separate questions:

• What is the central forecast?
• How uncertain is that forecast?

The ensemble mean answers the first question. Ensemble spread helps answer the second.

The ensemble mean is the average of all ensemble-member forecasts.

It is often smoother and more stable than an individual deterministic run because extreme member outcomes partially offset each other.

For HDD and CDD analysis, the ensemble mean provides a practical central estimate of expected weather demand.

However, the mean alone can be misleading.

A group of tightly clustered members and a widely dispersed group can produce the same average value. Their forecast confidence is not the same.

P10 is the 10th percentile of the ensemble distribution.

P90 is the 90th percentile.

After the ensemble values are ordered from lowest to highest:

• Approximately 10 percent of member outcomes fall below P10
• Approximately 90 percent fall below P90
• The range between P10 and P90 contains the middle 80 percent of ensemble-member outcomes

For degree days, the direction depends on the metric.

For HDD, higher values represent colder and more heating-demand-intensive outcomes.

For CDD, higher values represent hotter and more cooling-demand-intensive outcomes.

A narrow P10-P90 range means the ensemble members are relatively close together.

This suggests stronger agreement around the central forecast.

A wide P10-P90 range means the members show a broader set of possible outcomes.

This indicates greater forecast uncertainty.

The range normally becomes wider further into the forecast period because small differences between members grow over time.

The range should therefore be interpreted in the context of forecast lead time. A modest spread in the near term may be notable, while a wider range at the end of the forecast may be expected.

Consider two HDD forecasts.

Both forecasts have the same ensemble mean.

However, Forecast A has a narrow range, suggesting that most members support a similar demand outcome.

Forecast B includes much colder and much warmer possibilities. The central forecast is the same, but confidence around it is much lower.

Looking only at the mean would miss this difference.

Ensemble spread helps answer practical questions such as:

• Is a colder or hotter model shift supported by most members?
• Is the mean being influenced by a smaller group of extreme outcomes?
• Is forecast uncertainty increasing or decreasing?
• Are materially different demand scenarios still plausible?
• Do ensemble percentile ranges overlap across models?
• Is a forecast change becoming more stable across successive runs?

A large run-to-run change accompanied by narrowing spread may be more meaningful than the same change occurring while uncertainty continues to increase.

However, a narrow range does not guarantee accuracy. It only shows that the members in the current run agree more closely with each other.

P10-P90 ranges are also useful when comparing ensemble systems that publish percentile outcomes.

If the two means differ but their ranges overlap substantially, the disagreement may be less severe than the central values suggest.

If both the means and the ranges remain clearly separated, the models support meaningfully different demand scenarios.

Useful questions include:

• Do the ranges overlap?
• Is one model's entire range colder or hotter?
• Is one system showing much greater uncertainty?
• Are the ranges moving closer together in recent runs?
• Is the disagreement concentrated on specific dates?

Cross-model spread analysis provides more context than comparing only two mean lines.

A widening range indicates that the ensemble members are becoming more dispersed.

Possible interpretations include:

• A weather pattern is becoming less predictable
• The timing of a transition remains uncertain
• Different members are developing competing ridge or trough solutions
• Forecast confidence is declining

A widening range does not automatically mean the central forecast is wrong. It means a broader set of outcomes is being represented.

A narrowing range indicates increasing agreement among ensemble members.

This may occur as:

• The forecast period moves closer
• A weather pattern becomes better established
• Competing scenarios disappear from the ensemble
• Successive runs converge around a similar solution

A narrowing range can add credibility to a persistent HDD or CDD shift, especially when both GEFS and ECMWF EPS move in the same direction.

P10-P90 should not be treated as a guaranteed boundary.

Weather outcomes can still fall outside the range.

The range represents the distribution of ensemble members in one specific forecast run. It does not include every possible atmospheric outcome and is not a promise that the observed result will remain inside the band.

It also does not mean that every value inside the range is equally likely.

P10-P90 should be used as an uncertainty indicator, not as a precise trading probability.