measuring the performance of climate predictions chris ferro , tom fricker , david stephenson

17
Measuring the performance of climate predictions Chris Ferro , Tom Fricker, David Stephenson Mathematics Research Institute University of Exeter, UK MA Conference on the Mathematics of the Climate System eading, 14 September 2011

Upload: cleary

Post on 24-Feb-2016

57 views

Category:

Documents


0 download

DESCRIPTION

Measuring the performance of climate predictions Chris Ferro , Tom Fricker , David Stephenson Mathematics Research Institute University of Exeter, UK. IMA Conference on the Mathematics of the Climate System Reading, 14 September 2011. How good are climate predictions?. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Measuring theperformance ofclimate predictions

Chris Ferro, Tom Fricker,David Stephenson

Mathematics Research InstituteUniversity of Exeter, UK

IMA Conference on the Mathematics of the Climate SystemReading, 14 September 2011

Page 2: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

How good are climate predictions?

Predictions are useless without some information about their quality.

Focus on information contained in hindcasts, i.e. retrospective forecasts of past events.

1. How should we measure the performance of climate predictions?

2. What does past performance tell us about future performance?

Page 3: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Hindcasts

Thanks: Doug Smith (Met Office Hadley Centre)

Page 4: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Challenges

Sample sizes are small, e.g. CMIP5 core hindcast experiments give 10 predictions for each lead time.

Some external forcings (e.g. greenhouse gases and volcanoes) are prescribed, not predicted.

The quality of measurements of predictands varies over time and space.

Observations from the hindcast period are used (to some extent) to construct the prediction system.

Page 5: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Common practice

Choice of predictand:Evaluate predictions only after removing biasesEvaluate predictions of only long-term averages

Choice of performance measure:Evaluate only the ensemble mean predictionsEvaluate using correlation or mean square errorResample to estimate the sampling uncertainty

Page 6: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Common practice

Choice of predictand:Evaluate predictions only after removing biasesEvaluate predictions of only long-term averages

Choice of performance measure:Evaluate only the ensemble mean predictionsEvaluate using correlation or mean square errorResample to estimate the sampling uncertainty

Page 7: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Conventional reasoning

We can’t predict weather at long lead times.

So, don’t compare predicted and observed weather.

Instead, compare predicted and observed climate, e.g. multi-year averages.

Reduces noise and increases evaluation precision.

Page 8: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Evaluate weather, not climate!

The foregoing argument is wrong for two reasons.

We should evaluate predictands relevant to users.

Evaluating climate averages reduces signal-to-noise ratios and so decreases evaluation precision.

Better to evaluate predictions as weather forecasts and then average over time to improve precision.

Page 9: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Di prediction error for lead time i = 1, ..., nD error after averaging over the n lead timesS1 mean of the square of the errors D1, ..., Dn

Sn square of the mean error D

Under moderate conditions, the signal-to-noise ratio E(Sn)2 / var(Sn) of Sn becomes increasingly small relative to the signal-to-noise ratio of S1 as the averaging length, n, increases.

Evaluate weather, not climate!

Page 10: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Common practice

Choice of predictand:Evaluate predictions only after removing biasesEvaluate predictions of only long-term averages

Choice of performance measure:Evaluate only the ensemble mean predictionsEvaluate using correlation or mean square errorResample to estimate the sampling uncertainty

Page 11: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Skill inflation

Predictions initialized along trending observations.

Page 12: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Skill inflation

Strong association even if predictions fail to follow observations over the lead time.

Performance measures can mislead and mask differences between prediction systems.

Page 13: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Avoiding skill inflation

Observations Xt and predictions Pt sampled over time t from a joint distribution function F.

Real-valued performance measure, s(F).

Suppose that the joint distribution, Ft, of (Xt, Pt) changes with t so that F is a mixture distribution.

No skill inflation if s satisfies the following property:

s(Ft) = s0 for all t implies s(F) = s0 for all mixtures F.

Page 14: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Avoiding skill inflation

All convex properties of real-valued scoring rules, σ(X,P), are immune to skill inflation. These include

s(F) = expected value of σ(X,P),

e.g. mean square error, and

s(F) = any quantile of σ(X,P),

e.g. median absolute deviation.

Also monotonic functions of these, e.g. RMSE.

Page 15: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Summary

Measuring performance can help to improve predictions and to guide responses to predictions.

Evaluating climate predictions is hard because of small sample sizes, unpredicted forcings etc.

Evaluate as weather forecasts then average!

Use performance measures such as scoring rules that are immune to skill inflation from trends!

Page 16: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

Related questions

How does performance vary with the timescale of the predictand and of variations in the predictand?

What can we learn by evaluating across a range of lead times and evaluation periods?

What does past performance tell us about future performance?

How should hindcast experiments be designed to yield as much information as possible?

Page 17: Measuring the performance of climate predictions Chris Ferro , Tom  Fricker , David Stephenson

ReferencesFerro CAT, Fricker TE (2011) An unbiased decomposition of the Brier score.

Submitted.Fricker TE, Ferro CAT (2011) A framework for evaluating climate predictions. In

preparation.Goddard L and co-authors (2011) A verification framework for interannual-to-

decadal prediction experiments. In preparation.Jolliffe IT, Stephenson DB (2011) Forecast Verification: A Practitioner’s Guide in

Atmospheric Science. 2nd edition. Wiley. In press.Smith DM and co-authors (2007) Improved surface temperature prediction for

the coming decade from a global climate model. Science, 317, 796—799.

The EQUIP project: [email protected]