using probabilistic forecasts in decision makingwhy probabilistic forecasting? not only does the...

40
Using Probabilistic Forecasts in Decision Making Devin Kilminster Senior Scientist Forecasting Research Meteorological Service of New Zealand Limited NZWEA 2014 POWERFUL WEATHER INTELLIGENCE.

Upload: others

Post on 16-Sep-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Using Probabilistic Forecastsin Decision Making

Devin Kilminster

Senior ScientistForecasting Research

Meteorological Service of New Zealand Limited

NZWEA 2014

POWERFUL WEATHER INTELLIGENCE.

Page 2: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Why probabilistic forecasting?

• Not only does the forecasterpredict what is most likely tohappen,

• but they can also express theiruncertainty in this prediction.

Why should you care about the forecaster’s uncertainty?

My aims:

1. Present an example in which uncertainty information isvaluable.

2. Introduce the theory of decision making under uncertainty.

[Also:

• Contrast perspectives of the user vs. the provider of forecasts.]

POWERFUL WEATHER INTELLIGENCE.

Page 3: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Why probabilistic forecasting?

• Not only does the forecasterpredict what is most likely tohappen,

• but they can also express theiruncertainty in this prediction.

Why should you care about the forecaster’s uncertainty?

My aims:

1. Present an example in which uncertainty information isvaluable.

2. Introduce the theory of decision making under uncertainty.

[Also:

• Contrast perspectives of the user vs. the provider of forecasts.]

POWERFUL WEATHER INTELLIGENCE.

Page 4: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Why probabilistic forecasting?

• Not only does the forecasterpredict what is most likely tohappen,

• but they can also express theiruncertainty in this prediction.

Why should you care about the forecaster’s uncertainty?

My aims:

1. Present an example in which uncertainty information isvaluable.

2. Introduce the theory of decision making under uncertainty.

[Also:

• Contrast perspectives of the user vs. the provider of forecasts.]

POWERFUL WEATHER INTELLIGENCE.

Page 5: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Why probabilistic forecasting?

• Not only does the forecasterpredict what is most likely tohappen,

• but they can also express theiruncertainty in this prediction.

Why should you care about the forecaster’s uncertainty?

My aims:

1. Present an example in which uncertainty information isvaluable.

2. Introduce the theory of decision making under uncertainty.

[Also:

• Contrast perspectives of the user vs. the provider of forecasts.]

POWERFUL WEATHER INTELLIGENCE.

Page 6: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Why probabilistic forecasting?

• Not only does the forecasterpredict what is most likely tohappen,

• but they can also express theiruncertainty in this prediction.

Why should you care about the forecaster’s uncertainty?

My aims:

1. Present an example in which uncertainty information isvaluable.

2. Introduce the theory of decision making under uncertainty.

[Also:

• Contrast perspectives of the user vs. the provider of forecasts.]

POWERFUL WEATHER INTELLIGENCE.

Page 7: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Basics of pdfs

area under curve = probability

POWERFUL WEATHER INTELLIGENCE.

Page 8: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Basics of pdfs

less spread ⇒ more confidence

POWERFUL WEATHER INTELLIGENCE.

Page 9: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Basics of pdfs — levels and intervals

POWERFUL WEATHER INTELLIGENCE.

Page 10: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Basics of pdfs — range and exceedanceprobabilities

POWERFUL WEATHER INTELLIGENCE.

Page 11: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Example — Icing in CCGTs

When temperatures drop below 4C (humidity dependent), ice canform in CCGT turbine inlet region.

• Reduces efficiency (less power production).

• Increases wear and tear (increased maintenance costs).

• Potential for catastrophic damage (plant offline).

A de-icing agent can be applied:

• Best applied before ice begins to form (before T<4C):• less is needed to prevent than to remove ice.

• De-icing reduces profit by increasing operational costs.

POWERFUL WEATHER INTELLIGENCE.

Page 12: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Example — Icing in CCGTs

Simplifying our choices:

Tactic A ‘Wait and See’

• If T stays above 4 — no cost.• If T drops below 4 — high cost (eg $5000).

Tactic B ‘Pre-empt’

• Before T goes below 4 — low cost (eg $1000).• Chance we will be wrong and we have wasted

agent/money.

[These costs are purely hypothetical.]

Asymmetry of the costs suggests we want to be ‘conservative’.

POWERFUL WEATHER INTELLIGENCE.

Page 13: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

UK CCGT — 2 am early October 2012

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0

2

4

6

8

10

12

Tem

pera

ture

(C)

Current Observed: 4.9C

Current Forecast: 5.4C

Future Forecast: 4.9C

ObservedForecast

POWERFUL WEATHER INTELLIGENCE.

Page 14: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

UK CCGT — 2 am early October 2012

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0

2

4

6

8

10

12

Tem

pera

ture

(C)

Significant possibility of T<4C

ObservedForecast (50%, 90% intervals)

POWERFUL WEATHER INTELLIGENCE.

Page 15: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

UK CCGT — 2 am early October 2012

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0

2

4

6

8

10

12

Tem

pera

ture

(C)

Significant possibility of T<4C

Forecast uncertainty can decrease

ObservedForecast (50%, 90% intervals)

POWERFUL WEATHER INTELLIGENCE.

Page 16: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

UK CCGT — Hypothetical possibility

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0

2

4

6

8

10

12

Tem

pera

ture

(C)

Much less chance of T<4C

ObservedForecast (50%, 90% intervals)

POWERFUL WEATHER INTELLIGENCE.

Page 17: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Decision theory

1. Enumerate the possible actions.

2. For each action enumerate the relevant outcomes andconsequences.

3. Using the forecast, find the distributions of the consequencesfor each action.

4. Choose the action corresponding to the preferred distributionof consequences.

Action

Wait

Pre-empt

p is forecast probability that T< 4C.Pre-empt when 1000 < 5000p, ie when p > 20%.

POWERFUL WEATHER INTELLIGENCE.

Page 18: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Decision theory

1. Enumerate the possible actions.

2. For each action enumerate the relevant outcomes andconsequences.

3. Using the forecast, find the distributions of the consequencesfor each action.

4. Choose the action corresponding to the preferred distributionof consequences.

Action Outcome Consequence

Wait T< 4C Lose $5000T≥ 4C None

Pre-empt any Lose $1000

p is forecast probability that T< 4C.Pre-empt when 1000 < 5000p, ie when p > 20%.

POWERFUL WEATHER INTELLIGENCE.

Page 19: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Decision theory

1. Enumerate the possible actions.

2. For each action enumerate the relevant outcomes andconsequences.

3. Using the forecast, find the distributions of the consequencesfor each action.

4. Choose the action corresponding to the preferred distributionof consequences.

Action Outcome Consequence Probability

Wait T< 4C Lose $5000 pT≥ 4C None 1− p

Pre-empt any Lose $1000 1

p is forecast probability that T< 4C.

Pre-empt when 1000 < 5000p, ie when p > 20%.

POWERFUL WEATHER INTELLIGENCE.

Page 20: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Decision theory

1. Enumerate the possible actions.

2. For each action enumerate the relevant outcomes andconsequences.

3. Using the forecast, find the distributions of the consequencesfor each action.

4. Choose the action corresponding to the preferred distributionof consequences.

Action Outcome Consequence Probability E(Loss)

Wait T< 4C Lose $5000 p 5000pT≥ 4C None 1− p

Pre-empt any Lose $1000 1 1000

p is forecast probability that T< 4C.Pre-empt when 1000 < 5000p, ie when p > 20%.

POWERFUL WEATHER INTELLIGENCE.

Page 21: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Use Probability of Exceedance?

0

2

4

6

8

10

12Te

mpe

ratu

re (C

) ObservedForecast (50%, 90% intervals)

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0.000.050.100.150.200.250.300.350.40

Pr(T<

4)

POWERFUL WEATHER INTELLIGENCE.

Page 22: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Solving the problem ‘deterministically’

• Asymmetry of costs suggests choosing some ‘safety margin’:

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0

2

4

6

8

10

12

Tem

pera

ture

(C)

ObservedForecastForecast-SafetyMargin

• Really just estimating a constant forecast uncertainty.• Why not let the forecaster provide the uncertainty?

• Sub-optimal if forecast uncertainty is actually non-constant.

POWERFUL WEATHER INTELLIGENCE.

Page 23: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Solving the problem ‘deterministically’

• Asymmetry of costs suggests choosing some ‘safety margin’:

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0

2

4

6

8

10

12

Tem

pera

ture

(C)

ObservedForecastForecast-SafetyMargin

• Really just estimating a constant forecast uncertainty.• Why not let the forecaster provide the uncertainty?

• Sub-optimal if forecast uncertainty is actually non-constant.

POWERFUL WEATHER INTELLIGENCE.

Page 24: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Solving the problem ‘deterministically’

• Asymmetry of costs suggests choosing some ‘safety margin’:

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0

2

4

6

8

10

12

Tem

pera

ture

(C)

ObservedForecastForecast-SafetyMargin

• Really just estimating a constant forecast uncertainty.• Why not let the forecaster provide the uncertainty?

• Sub-optimal if forecast uncertainty is actually non-constant.

POWERFUL WEATHER INTELLIGENCE.

Page 25: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Some other issues

1. Using hourly pdfforecasts, but really careabout extreme across arange of time.

2. Information inobservations made afterforecast was issued.

0

2

4

6

8

10

12

Tem

pera

ture

(C) Observed

Forecast (50%, 90% intervals)

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0.000.050.100.150.200.250.300.350.40

Pr(T<

4)

1. Joint uncertainty information.

• Use information about correlations in observations wrtforecasts — high in this case.

• Forecast the quantity directly relevant to the decision.(e.g. the minimum.)

2. Frequent updating of forecast.• Provide real-time observations to the forecaster?

POWERFUL WEATHER INTELLIGENCE.

Page 26: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Some other issues

1. Using hourly pdfforecasts, but really careabout extreme across arange of time.

2. Information inobservations made afterforecast was issued.

0

2

4

6

8

10

12

Tem

pera

ture

(C) Observed

Forecast (50%, 90% intervals)

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0.000.050.100.150.200.250.300.350.40

Pr(T<

4)

1. Joint uncertainty information.

• Use information about correlations in observations wrtforecasts — high in this case.

• Forecast the quantity directly relevant to the decision.(e.g. the minimum.)

2. Frequent updating of forecast.• Provide real-time observations to the forecaster?

POWERFUL WEATHER INTELLIGENCE.

Page 27: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Some other issues

1. Using hourly pdfforecasts, but really careabout extreme across arange of time.

2. Information inobservations made afterforecast was issued.

0

2

4

6

8

10

12

Tem

pera

ture

(C) Observed

Forecast (50%, 90% intervals)

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0.000.050.100.150.200.250.300.350.40

Pr(T<

4)

1. Joint uncertainty information.• Use information about correlations in observations wrt

forecasts — high in this case.

• Forecast the quantity directly relevant to the decision.(e.g. the minimum.)

2. Frequent updating of forecast.• Provide real-time observations to the forecaster?

POWERFUL WEATHER INTELLIGENCE.

Page 28: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Some other issues

1. Using hourly pdfforecasts, but really careabout extreme across arange of time.

2. Information inobservations made afterforecast was issued.

0

2

4

6

8

10

12

Tem

pera

ture

(C) Observed

Forecast (50%, 90% intervals)

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0.000.050.100.150.200.250.300.350.40

Pr(T<

4)

1. Joint uncertainty information.• Use information about correlations in observations wrt

forecasts — high in this case.• Forecast the quantity directly relevant to the decision.

(e.g. the minimum.)

2. Frequent updating of forecast.• Provide real-time observations to the forecaster?

POWERFUL WEATHER INTELLIGENCE.

Page 29: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Some other issues

1. Using hourly pdfforecasts, but really careabout extreme across arange of time.

2. Information inobservations made afterforecast was issued.

0

2

4

6

8

10

12

Tem

pera

ture

(C) Observed

Forecast (50%, 90% intervals)

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0.000.050.100.150.200.250.300.350.40

Pr(T<

4)

1. Joint uncertainty information.• Use information about correlations in observations wrt

forecasts — high in this case.• Forecast the quantity directly relevant to the decision.

(e.g. the minimum.)

2. Frequent updating of forecast.• Provide real-time observations to the forecaster?

POWERFUL WEATHER INTELLIGENCE.

Page 30: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

When does decision theory work?

Reliability:1

p(·|f ) = f

• Roughly, that no ‘recalibration’ of the forecast is necessary toget probabilities.

• Implies, for example, ‘flatness’ of the Rank Histogram:

0.0 0.2 0.4 0.6 0.8 1.0Forecast Probability T<4C

0.00

0.05

0.10

0.15

0.20

0.25

0.30

Obse

rved

Fre

quen

cy T<

4C

Rank Histogram

ideal2σ

• Often used as test ofreliability — but maybesomething more relevantto the decision isappropriate. . . .

1see for example Scoring Probabilistic Forecasts: The Importance of Being Proper, J. Brocker and L. Smith,

Weather and Forecasting, 22, 2007.

POWERFUL WEATHER INTELLIGENCE.

Page 31: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

When does decision theory work?

Reliability:1

p(·|f ) = f

• Roughly, that no ‘recalibration’ of the forecast is necessary toget probabilities.

• Implies, for example, ‘flatness’ of the Rank Histogram:

0.0 0.2 0.4 0.6 0.8 1.0Forecast Probability T<4C

0.00

0.05

0.10

0.15

0.20

0.25

0.30

Obse

rved

Fre

quen

cy T<

4C

Rank Histogram

ideal2σ

• Often used as test ofreliability — but maybesomething more relevantto the decision isappropriate. . . .

1see for example Scoring Probabilistic Forecasts: The Importance of Being Proper, J. Brocker and L. Smith,

Weather and Forecasting, 22, 2007.

POWERFUL WEATHER INTELLIGENCE.

Page 32: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

When does decision theory work?

Reliability:1

p(·|f ) = f

• Roughly, that no ‘recalibration’ of the forecast is necessary toget probabilities.

• Implies, for example, ‘flatness’ of the Rank Histogram:

0.0 0.2 0.4 0.6 0.8 1.0Forecast Probability T<4C

0.00

0.05

0.10

0.15

0.20

0.25

0.30

Obse

rved

Fre

quen

cy T<

4C

Rank Histogram

ideal2σ

• Often used as test ofreliability — but maybesomething more relevantto the decision isappropriate. . . .

1see for example Scoring Probabilistic Forecasts: The Importance of Being Proper, J. Brocker and L. Smith,

Weather and Forecasting, 22, 2007.

POWERFUL WEATHER INTELLIGENCE.

Page 33: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

0.0 0.2 0.4 0.6 0.8 1.0Threshold probability T<4C

10

5

0

5

10

15

20

25

30

% c

ost v

s th

eore

tical

opt

imal

thre

shol

d

Reliability of Forecasts for CCGT de-icing problem

1 σ

My goal as a forecast provider is to make probabilistic forecaststhat:• are reliable with respect to my customer’s decision problems,

and• are accurate in the sense that they can help my customer

make good decisions.

POWERFUL WEATHER INTELLIGENCE.

Page 34: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

0.0 0.2 0.4 0.6 0.8 1.0Threshold probability T<4C

10

5

0

5

10

15

20

25

30

% c

ost v

s th

eore

tical

opt

imal

thre

shol

d

Reliability of Forecasts for CCGT de-icing problem

1 σ

My goal as a forecast provider is to make probabilistic forecaststhat:• are reliable with respect to my customer’s decision problems,

and• are accurate in the sense that they can help my customer

make good decisions.

POWERFUL WEATHER INTELLIGENCE.

Page 35: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Summary

• Probabilities allow forecasters to convey extra (uncertainty)information.

• Information about uncertainty can allow better decisions.

• Decision theory.2

• Are we forecasting the quantity most relevant to the decision?— Forecast of Minimum vs. Hourlies.

• Reliability and Accuracy.

2For other examples see http://www.icem2011.org/presentations2011/2_Tuesday/1B/1345_Peters.pdf or

In Search of the Best Possible Weather Forecast for the Energy Industry, P. Mailier et. al. in Weather Matters forEnergy, Springer 2014.

POWERFUL WEATHER INTELLIGENCE.

Page 36: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Summary

• Probabilities allow forecasters to convey extra (uncertainty)information.

• Information about uncertainty can allow better decisions.

• Decision theory.2

• Are we forecasting the quantity most relevant to the decision?— Forecast of Minimum vs. Hourlies.

• Reliability and Accuracy.

2For other examples see http://www.icem2011.org/presentations2011/2_Tuesday/1B/1345_Peters.pdf or

In Search of the Best Possible Weather Forecast for the Energy Industry, P. Mailier et. al. in Weather Matters forEnergy, Springer 2014.

POWERFUL WEATHER INTELLIGENCE.

Page 37: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Summary

• Probabilities allow forecasters to convey extra (uncertainty)information.

• Information about uncertainty can allow better decisions.

• Decision theory.2

• Are we forecasting the quantity most relevant to the decision?— Forecast of Minimum vs. Hourlies.

• Reliability and Accuracy.

2For other examples see http://www.icem2011.org/presentations2011/2_Tuesday/1B/1345_Peters.pdf or

In Search of the Best Possible Weather Forecast for the Energy Industry, P. Mailier et. al. in Weather Matters forEnergy, Springer 2014.

POWERFUL WEATHER INTELLIGENCE.

Page 38: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Summary

• Probabilities allow forecasters to convey extra (uncertainty)information.

• Information about uncertainty can allow better decisions.

• Decision theory.2

• Are we forecasting the quantity most relevant to the decision?— Forecast of Minimum vs. Hourlies.

• Reliability and Accuracy.

2For other examples see http://www.icem2011.org/presentations2011/2_Tuesday/1B/1345_Peters.pdf or

In Search of the Best Possible Weather Forecast for the Energy Industry, P. Mailier et. al. in Weather Matters forEnergy, Springer 2014.

POWERFUL WEATHER INTELLIGENCE.

Page 39: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

Summary

• Probabilities allow forecasters to convey extra (uncertainty)information.

• Information about uncertainty can allow better decisions.

• Decision theory.2

• Are we forecasting the quantity most relevant to the decision?— Forecast of Minimum vs. Hourlies.

• Reliability and Accuracy.

2For other examples see http://www.icem2011.org/presentations2011/2_Tuesday/1B/1345_Peters.pdf or

In Search of the Best Possible Weather Forecast for the Energy Industry, P. Mailier et. al. in Weather Matters forEnergy, Springer 2014.

POWERFUL WEATHER INTELLIGENCE.

Page 40: Using Probabilistic Forecasts in Decision MakingWhy probabilistic forecasting? Not only does the forecaster predict what ismost likelyto happen, but they can also express their uncertaintyin

What happened?

12Z 14Z 16Z 18Z 20Z 22Z 00Z 02Z 04Z 06Z 08Z 10Z 12ZTime

0

2

4

6

8

10

12

Tem

pera

ture

(C)

ObservedForecast (50%, 90% intervals)

POWERFUL WEATHER INTELLIGENCE.