1 wmo swfdp macau 9 april 2013 anders persson decision making process and blending ensemble and...

Post on 23-Dec-2015

216 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

1 WMO SWFDP Macau 9 April

2013 Anders Persson

Decision making process and blending ensemble and deterministic forecasts

2 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

1. What do good forecasters do?

3 WMO SWFDP Macau 9 April

2013 Anders Persson

Blending deterministic and probabilistic forecast information has been a challenge since Fitzroy started weather forecasting 150 years ago

An overcast evening outside London in January 1863:

Low clouds over snow covered ground with +2° C

The clouds will disperse and the temperature drop to -6° C

But will the clouds disperse?

Probably? (=60% chance?)

How will this affect the forecast?

4 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

The same situation in our days:

The NWP tells the clouds will clear and the temperature drop

+2º -6º

A classical, physical-meteorological, deterministic problemThe skilled weather forecaster is invited to “add value” to the NWP by modifying the -6° forecast

However, the real added value might be of some other kind. . . assume the clouds do not lift?

5 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

Assume the probability of clearing = 60% Three different forecasts might be provided

I. A compromise forecast -3º for verifications

II. A missed event is considered worse than a false alarm so -4º or -5º is forecast

III. Special customers are told that there is a slightly higher probability (60%) for the clouds to disperse with -6º, rather than not (40%) with +2º

All of these involve clever use of intuitive statistics

6 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

Asymmetric cost- or penalty functions

Error Error

“Pain”

Possible errors Possible errors

● ●

“Pain”Forecasts for unspecified

customer or for verification purposes

Specific customer

with sensitivity for missed cold

events

Forecast

too mild

Forecast too cold

7 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

Expected mean cost/d

£30

£20

£10

£0

Value of probabilities : The school book example

Loss=£100 and average probability of bad weather pclim=30%

£0 £30 £60 £90 protection cost

£30

£20

£10

0

gain

Never protect

Always protect

Deterministic forecast

Perfect forecasts

Useful

forecasts

Ob Fc R _

R 20 10 - 10 60

8 WMO SWFDP Macau 9 April

2013 Anders Persson

Ob Fc R _

R 20 10 - 10 60

ObProb R _

100 10 0 80 8 2 60 6 4 40 4 6 20 2 8 0 0 50

Ob Fc R -

R 10 0

?? 20 20

- 0 50

Categorical Non-categorical

The value of uncertain weather forecasts

Probabilistic

9 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

Loss=£100 and average probability of bad weather pclim=30%

gain

Never protect

Always protect

Deterministic forecast

Probabilistic forecasts

Ob%

R _

100 10 0

80 8 2

60 6 4

40 4 6

20 2 8

0 0 50

Expected mean cost/d

£30

£20

£10

£0£0 £30 £60 £90 protection cost

£30

£20

£10

0

10 WMO SWFDP Macau 9 April

2013 Anders Persson

The intuitive-statistical nature of routine forecasting

The forecasters work in an environment with a flow of information from different sources that might be incorrect, contradictory and unrepresentative

11 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

2. The need of good “statistical intuition” has been the subject of learned books

12 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

Not only meteorologists are concerned with risks and uncertainties

13 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

Time constrains, limited and sometimes misleading information, stress and outside distraction

(Almost) unlimited time, a wide range of reliable information

and full concentration

Fast thinking: Meteorologists in the forecast office

Slow thinking: Meteorologists attending a seminar

The title of Kahneman´s book refers to

14 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

3. Five points where we humans have to improve on how to deal with uncertainties

15 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23 15

Common human intuitive weaknesses

1. Over-confidence

2. Underestimation of randomness

3. Problems estimating uncertainty

4. Communicating this uncertainty

5. Drawing the conclusions from uncertainty

16 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

3.1 Overconfidence

17 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/2319/04/23 17

3.1 Overconfidence: Before 2000 Concord was regarded as the safest airplane

Concorde Other company

__0___ 100 000Flight hours

< __1_____ 1 000 000Flight hours

__1___ 100 001Flight hours

> __1_____ 1 000 000Flight hours

. . .after the 2000 crash the most unsafe

accidents

accidents

18 WMO SWFDP Macau 9 April

2013 Anders Persson

⎟ ⎟ ⎟Mon 00 Mon 12 Tue 00 Tue 12 Wed 00 Wed 12

Model 1

Model 2

Model 3

Model 1

Model 2

Model 3

⎟ ⎟ ⎟

Mon 00 Mon 12 Tue 00 Tue 12 Wed 00 Wed 12

Over confidenceThree forecasts from different NWP models valid at the same time

in 5% of the cases

Surely dry!

in 80% of the cases

Surely rain!

19 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

3.2.Underestimating randomness

20 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/2319/04/23 20

Conditional sampling

Comments from ECMWF Member States:

1. You overforecast Portuguese cut-offs at D+5, only 50% verify

2. You overforecast >25 mm/day events at D+3, only 50% verify

3. You overforecast gales at D+4, only 50% verify

OBFC

OB OB

FCFC

Many hitsUnder forecasting

Few missesOver forecasting

Well tuned forecasts

21 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

3.3 Estimating uncertainty (probabilities)

22 WMO SWFDP Macau 9 April

2013 Anders Persson

⎟ ⎟ ⎟Mon 00 Mon 12 Tue 00 Tue 12 Wed 00 Wed 12

Forecast from model A

Forecast from model B

Forecast from model C

The Halo Effect

“…the atmosphere is inherently unpredictable due to the chaotic nature of its motions (Ørgård, 1963)…”

“…the atmosphere is inherently unpredictable due to the chaotic nature of its motions (Lorenz, 1963)…”

In forecasting meteorology: to weight one’s favourite NWP too much

23 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23Barcelona 5 Sep 2012

Anders Persson

15 UTC chart 15 UTC forecast

03 UTC chart

(≡) =

=15 UTC forecast

☼Δ

Δ

Δ

Δ+24 h forecast ☼☼

The availability effect

( )( )

TS-risk 70%

TS-risk 30%

+12 h forecast

24 WMO SWFDP Macau 9 April

2013 Anders Persson

⎟ ⎟ ⎟Mon 00 Mon 12 Tue 00 Tue 12 Wed 00 Wed 12

Model 1

Model 2

Model 3

Model 1

Model 2

Model 3

⎟ ⎟ ⎟

Mon 00 Mon 12 Tue 00 Tue 12 Wed 00 Wed 12

The primacy effectTo order of arrival of the NWP may also affect the assessment

“Yes, rain is possible”

“No, I do not believe in rain”

25 WMO SWFDP Macau 9 April

2013 Anders Persson

in 63% of the cases

⎟ ⎟ ⎟Mon 00 Mon 12 Tue 00 Tue 12 Wed 00 Wed 12

Forecast 1

Forecast 2

Forecast 3

Forecast 1

Forecast 2

Forecast 3

⎟ ⎟ ⎟

Mon 00 Mon 12 Tue 00 Tue 12 Wed 00 Wed 12

Misleading consistencyThree consecutive NWP from the same model valid at the same time

in 58% of the cases

Very “jumpy”

Rather consistent

26 WMO SWFDP Macau 9 April

2013 Anders Persson

Consecutive forecasts tend to be correlated since the new observations do not change the “first guess” entirely

The two best on average but also the most correlated ones i.e. their

mutual agreement is less significant

The best and the worst on average but also the least correlated ones. Their mutual agreement becomes more significant

+24h

+36h

+48h

+24h

+36h

+48h

27 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

3.4 Communicating uncertainty

28 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23 28

People react differently to a statement like:

“-There is a 30% risk of rain”

compared to

”- A 70% chance of dry weather”

This is the “Framing effect”

3.4 Example of communicating probabilities

29 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23 29

An example of a meteorological framing effect:

The authorities react more appropriately to a probability forecast of 60% for a whole region (Midlands) than to 10-20% for an individual location (Birmingham)

20%60%15%

10%

Thunderstorm risk

30 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/2319/04/23 30

The Base rate effect

50% probability means different things Base rate

1. Tossing a coin: 50-50? = I do not know 50%

2. Snowfall in Barcelona: 50% very high risk! 2%

3. <4/8 clouds in Barcelona: 50% is a low “risk”! 80%

It all depends on the “base rate”

31 WMO SWFDP Macau 9 April

2013 Anders Persson

The base rate in meteorology is the climatology

The ECMWF:s Extreme Forecast Index (EFI) relates the probabilities to the climatology

19/04/23 31

ECMWF’s new EFI chart 16 August 2012 12 UTC +60 to +84 h

32 WMO SWFDP Macau 9 April

2013 Anders Persson

Excessive rain risk

Excessive hot

33 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

D+8 forecast 7 December

D+7 forecast 8 December

D+6 forecast 9 December

D+5 forecast 10 December

The predicted arrivals of the 15-16 December storm(ECMWF and UKMO alike)

34 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

The cyclone has changed track several times - we have revised our

calculations

No blame on the computer for the “jumpiness”

35 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/2319/04/23 35

The way the Met Office and BBC forecasters handled the weather situation was “very well received by senior managers in the BBC and the Met Office….and had been praised by the section of government which is responsible for the Met Office. “

No direct surveys of public opinion were made, “but informal feedback has been positive.”

I’ll come back into more detail what the BBC/Met Office did

36 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

3.5 Drawing conclusions from probabilities

-What do you prefer?

-An 80% chance of winning £1000 or

-Get £700 directly?

37 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

Summary of part 1-3:

• The forecasters will increasingly deal with forecast uncertainty and risk assessments, which will increase the public’s confidence and improve the weather forecasters´ reputation

• A five-point program is suggested on how to change the current deterministic culture:

• The greatest “threat” to the meteorological weather forecaster is not the computer but the growing number of non-meteorological weather forecasters with a modern outlook

a) Reduce forecast over-confidenceb) Understand the effects of randomnessc) To estimate forecast uncertainty d) To convey probabilistic informatione) To help the customers to make decisions

38 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

4. Updating deterministic and ensemble forecasts

39 WMO SWFDP Macau 9 April

2013 Anders Persson

Meteorologists have five sources of information:

1. Observations

2. Deterministic NWP

3. Statistical interpretation

4. Ensemble forecasts

5. Climatological information

Let’s start with the last one, point 5.

(systematic errors and “jumpiness”)

(irregular, varying quality and representative)

(outdated or unrepresentative)

(a lot of information with “probs”)

(“only” background information)

40 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

4.1 The problem seen from a typical PDF (probability density function) perspective – climate distribution

41 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ2% 14% 34% 34% 14% 2%

Mean value+1 SD +2 SD-1 SD-2 SD

Most likely

valuesHigher than normal

Much higher than normal

Lower than normal

Much lower than

normal

A typical climatological distribution(temperature)

42 WMO SWFDP Macau 9 April

2013 Anders Persson

very tricky for bi-modal distributions

More tricky for rain or wind

Mode (the most likely value) Median (divides the data into two equal halves)

Mean (the average)

Mode

MedianMean

Mode

Median

Mean

What is one single “representative” value?

43 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Climatological average

Probability

The analysisobservations

The forecast

44 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Larger area = more certain

Smaller area = less certain

If the forecast is wrong it is more likely to be wrong “to the left” (less anomalous) than “to the right” (even more anomalous)

The “Regression to the Mean

Effect”

Probability

45 WMO SWFDP Macau 9 April

2013 Anders Persson

ΨClimatological average

More probable

Less probable

Probability

46 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

How certain is this NWP?

We do not know!

It could be very certain

. . . or very uncertain

NWP

47 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

Investigations show that “jumpiness” correlates badly to the accuracy of the last forecast

NWP today

NWP yesterdayNWP the day

before yesterday

We might an opinion by looking at the last NWPs from the same model, so called “lagged” forecasts

48 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

The “jumpiness” correlates fairly well to the accuracy of the weighted average of the three NWP

NWP today

NWP yesterdayNWP the day

before yesterday

This “poor man’s ensemble captures the essentials of the “rich man’s” ensemble

1. Ensemble mean2. Spread3. Rough “probs”

Weighted average of the three

NWP

49 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

Consensus NWP

There are no exact rules on how to merge manual and NWP information

Subjectively weighted average of the

manual and NWP

ManualFinal?

50 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

Then arrives the EPS, the

Ensemble Prediction System forecast

FinalAgain, there are no exact rules on how

to merge manual and

NWP forecasts

51 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

Again, there are no exact rules on how to merge manual and EPS information

EPS

Manual+NWP

New

52 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

The change in the most representative value, from the manual+NWP to the new forecast, is not much affected

Manual+NWP

EPSNew

Minor change of deterministic value

53 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

The major change is in the spread of the forecasts, the (un)certainty

Major change of probabilistic values

Increased risk for extremes

54 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

4.2 The problem seen from an EPS meteogram perspective

55 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

Medium range forecasting…with deterministic and EPS information

Latest three NWP

Latest EPS

Most common case with good agreement between EPS spread

and NWP “jumpiness”

Most common case

56 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

Rather poor agreement between larger EPS

spread and small NWP “jumpiness”.

The analysis system has obviously managed to

avoid possible problems because the NWP is not

very “jumpy”

Should the forecasters be more certain than the EPS indicates?

Rather common case

57 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

Rather poor agreement between small EPS

spread and large NWP “jumpiness”.

The perturbations have not been quite

able to cover the analysis uncertainties

Should the forecasters be more uncertain than the EPS indicates?

Not uncommon case

58 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

Poor agreement between the main

directions of the EPS and the NWP

This puts the forecasters in a very difficult situation and there is not enough experience or investigations about this situation

Rare case

Best choice: create a “super ensemble”

59 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

4.3 The same seen from a PDF-perspective

60 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

Again, there are no exact rules on how to merge manual and EPS information

EPS

Manual+NWP

New

61 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

EPS

NWP

Final

Case 1: Lagged NWP agree with EPS and about the same spread

62 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

EPS

NWP

Final

Case 2: Lagged NWP agree with EPS but has smaller spread

63 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

EPS

NWP

Final

Case 3: Lagged NWP agree with EPS but has larger spread

64 WMO SWFDP Macau 9 April

2013 Anders Persson

Ψ

Probability

EPS

NWP

Final

Case 4: Lagged NWP agree with EPS and about the same spread but quite different means

65 WMO SWFDP Macau 9 April

2013 Anders Persson

Summary of part 4

1. The forecaster has an increasing role to play as an “intuitive statistician”

2. The EPS must be compared and blended with more than one NWP, preferably 3-4 NWP

3. In the blending the spread and probabilities will normally be affected more than the “representative” value.

Ensemble information tend to make us more uncertain – is that good?

66 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

5. The value of uncertainty “per se”

67 WMO SWFDP Macau 9 April

2013 Anders Persson

Ob Fc R _

R 20 10 - 10 60

Ob Fc R -

R 10 0

?? 20 20

- 0 50

Categorical Non-categorical

The value of uncertain weather forecasts

Ob Fc R _

R 30 20 - 0 50

Ob Fc R _

R 10 0 - 20 70

Low protection cost

High protection cost

68 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

Loss=£100 and average probability of bad weather pclim=30%

gain

Ob Fc R _

R 20 10- 10 60

Low protection

costHigh protection cost

Ob Fc R -

R 20 10

?? 20 20

- 10 60

Expected mean cost/d

£30

£20

£10

£0£0 £30 £60 £90 protection cost

£30

£20

£10

0

This is not just playing with mathematics – this

was the approach actually used by the Met

Office and BBC in December 2011

69 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

“Some terrible weather will come on Thursday-Friday”

The BBC forecasters avoided going into

detail and did not show any isobar

maps

70 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

On Wednesday 14 Dec still large uncertainty about the storm track

71 WMO SWFDP Macau 9 April

2013 Anders Persson

. . . and then finally the day before

72 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

The Met Office repeated the approach 1 ½ month later

73 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/23

The Met Office and the BBC didn’t hide, but made use of the uncertainty

74 WMO SWFDP Macau 9 April

2013 Anders Persson

My conclusions from the Met Office and BBC experience

1. Uncertainties can be communicated without numbers

2. The meteorologist must appear to be in control

3. Tell the background to the uncertainty – tell a “story”

4. Do not hesitate to give advice such as “if I were you…”

5. Follow up the forecast – but do not take to much credit

75 WMO SWFDP Macau 9 April

2013 Anders Persson19/04/2319/04/23 75

End

76 WMO SWFDP Macau 9 April

2013 Anders Persson

Observed rain 9-12 July 2004

50 mm

Prognosis

8 July

30 mm30 mmPrognosisPrognosis

7 July7 July

40 mmPrognosis

5 July

30 mmPrognosis

6 July

Expected rain for 9 July 2004

Example from Sweden

The meteorologists’ forecasts

77 WMO SWFDP Macau 9 April

2013 Anders Persson

Observed rain 9-12 July 2004

Somewhere 50 mm

Somewhere 30 mmSomewhere 30 mmSomewhere 30 mm

Somewhere 30 mm

Expected rain for 9 July 2004

The hydrologists’ forecasts

78 WMO SWFDP Macau 9 April

2013 Anders Persson

top related