Weatherforecast Decision

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Meteorol. Appl. 9, 307–315 (2002) DOI:10.

1017/S1350482702003043

Decision-making from probability forecasts based


on forecast value
Kenneth R. Mylne The Met Office, London Road, Bracknell, Berks RG12 2SZ UK

A method of estimating the economic value of weather forecasts for decision-making is described. This
method has recently been used for user-oriented verification of probability forecasts, but is here applied
to aid forecast users in optimising their decision-making from probability forecasts. Value may be
calculated in the same way for either probability forecasts or deterministic forecasts, and thus provides
the user with a direct comparison of the value of each in terms of money saved, which is more relevant
than most standard verification scores. The method is illustrated using site-specific probability forecasts
generated from the ECMWF ensemble prediction system and deterministic forecasts from the ECMWF
high-resolution global model. It is found that for most forecast events and for most users the probability
forecasts have greater value than the deterministic forecasts from a higher resolution model.

1. Introduction Forecasts (ECMWF) (Molteni et al. 1996; Buizza and


Palmer 1998) and the US National Centers for
A weather forecast, however skilful, has no intrinsic Environmental Prediction (NCEP) (Toth and Kalnay
value unless it can be used to make decisions that bring 1993). Output from an EPS is normally in the form of
some benefit to the end user. Conventionally, a probability forecasts, and there is growing evidence
weather forecast provider supplies the user with a best (e.g. Molteni et al. 1996; Toth et al. 1997) that these
estimate of whether a defined event will occur (e.g. have greater skill than equivalent deterministic fore-
whether wind speed will or will not exceed 15 ms–1), or casts based on single high-resolution model forecasts,
of a value for a measurable parameter (e.g. maximum particularly on medium range time-scales. To make use
wind speed = 18 ms–1). Decision-making is often based of the additional information content of probability
on whether a defined event is expected to occur or not. forecasts, the decision-maker needs to know how to
For example, the owner of a small boat may decide to respond to a forecast such as ‘There is a 30% probabil-
seek shelter when the forecast wind speed exceeds ity that the wind speed will exceed 15 ms–1.’ This paper
15 ms–1. describes a technique which estimates the economic
value of a probability forecast system based on verifi-
The nature of atmospheric predictability is such that cation of past performance, and uses it to determine a
there is frequently a significant uncertainty associated user’s best decision-making strategy. The value of
with such deterministic (best estimate) forecasts. deterministic forecasts can be calculated in the same
Forecast uncertainty can be expressed in many ways, way, and this allows a direct comparison of the utility
either qualitatively or quantitatively, and where pro- of probability and equivalent deterministic forecasts in
vided this information can aid the decision-maker who terms which are relevant to the user. The method is
understands the impact of a wrong decision. However, illustrated using probability forecasts generated from
uncertainty is most commonly estimated subjectively; the ECMWF EPS and comparing them with forecasts
such estimates are often inconsistent, and may be from the ECMWF high-resolution deterministic
affected by factors such as the forecaster ‘erring on the model.
safe side’, which may not lead to optimal decision-mak-
ing for specific users. Over the years several objective
methods of generating probability forecasts have been 2. Background to probability forecasts based on
developed to take account of uncertainty. Probability ensembles
forecasts have long been generated using MOS (model
output statistics) where a deterministic model forecast Uncertainty in numerical weather forecasts derives
is augmented with a probability distribution based on a from a number of sources, in particular uncertainty in
previously observed distribution of model errors (e.g. the analysis of the initial state of the atmosphere and
Glahn and Lowry 1972; Carter et al. 1989). More approximations in the model used to predict the future
recently there has been rapid development of ensemble evolution. Errors in the analysis result from observa-
prediction systems (EPS) such as those operated by the tional errors, shortage of observations in some regions
European Centre for Medium Range Weather of the globe, and limitations of the data assimilation

307
14698080, 2002, 3, Downloaded from https://rmets.onlinelibrary.wiley.com/doi/10.1017/S1350482702003043 by Readcube (Labtiva Inc.), Wiley Online Library on [15/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Kenneth R. Mylne
system. Model errors are due to essential approxima- related to ROC verification, and is based on the cost-
tions in the formulation of a model, notably the many loss situation described by Murphy (1977). It has
small-scale processes which cannot be resolved explic- recently been discussed by Richardson (2000), and
itly, and whose effect must be approximated by para- is rapidly becoming accepted as a valuable tool for
meterisation. The non-linear nature of atmospheric user-oriented verification of probability forecasts
evolution means that small errors in the model repre- (e.g. Thornes and Stephenson 2001). The method
sentation of the atmospheric state, whether due to the has also been applied to seasonal forecasts by Graham
analysis or the model, will be amplified through the et al. (2000). The aim of this paper is to present the
course of a forecast and can result in large errors in the method in a way in which forecast providers and users,
forecast. This sensitivity was first recognised by working together, can optimise decision-making even
Lorenz (1963), and was influential in the development where the forecast probabilities may be significantly in
of chaos theory (Lorenz 1993). Gross errors in synop- error.
tic-scale evolution are common in medium-range fore-
casts (over 3 days), but can occasionally occur at less The concept behind forecast value is that forecasts only
than 24 hours. Errors in the detail of a forecast, such as have value if a user takes action as a result, and the
precipitation amounts or locations, are common even action saves the user money. Calculation of value for
in short-range forecasts. Ensemble prediction systems predictions of a defined event therefore requires infor-
have been developed in an attempt to estimate the mation on (a) the ability of the forecast system to pre-
probability density function (pdf) of forecast solutions dict the event, and (b) the user’s costs and losses asso-
by sampling the uncertainty in the analysis and running ciated with the various possible forecast outcomes.
a number of forecasts from perturbed analyses Consequently, the value depends on the application as
(Molteni et al. 1996; Toth and Kalnay 1993). In more well as on the skill of the forecast.
recent developments, Buizza et al. (1999) have included
some allowance for model errors in the ECMWF EPS,
by adding stochastic perturbations to the effects of 3.1 Ability of the forecast system
model physics. Houtekamer et al. (1996) describe an
ensemble which accounts for both model errors and ROC verification and the cost-loss situation are based
analysis errors by using a range of perturbations in on categorical forecasts of a binary event, against which
both model formulation and analysis cycles. With any a user can take protective action when the event is
of these ensemble systems, probability forecasts may be expected to occur. For such an event, the ability of a
generated by interpreting the proportion of ensemble deterministic forecast system is fully specified by the 22
members predicting an event as an estimate of the contingency table shown in Table 1, where h, m, f and
probability of that event. r give the relative frequencies of occurrence of each
possible forecast outcome.
A range of standard verification diagnostics are used to
assess the skill of probability forecasts. For example, Table 1. Contingency table of forecast performance.
the Brier Score (see Wilks 1995), originally introduced Upper-case letters H, M, F, R represent the total
by Brier (1950), is essentially the mean square error for numbers of occurrences of each contingency, while the
probability forecasts of an event. ROC (relative operat- lower-case letters in brackets represent the relative
ing characteristics), described by Stanski et al. (1989), frequencies of occurrences.
measures the skill of a forecast in predicting an event in
terms of hit rates and false alarm rates. Rank histograms Event Forecast
(Hamill and Colucci 1997, 1998) measure the extent to Yes – User No – No
which the ensemble spread covers the forecast uncer- Protects Protective
tainty, and can also reveal biases in the ensemble fore- Action
casts. However, while these diagnostics are useful to
scientists developing ensemble systems, they are of lit- Yes Hit H Miss M
tle relevance to most forecast users. They do not tell (h) (m)
users how valuable the forecasts will be for their appli- Events
No False Correct
cations, nor do they answer the question of how to use Observed
Alarm F Rejection R
probability forecasts for decision-making. (f) (r)

3. Cost-loss models and economic value of


forecasts 3.2 User costs and losses
An overview of techniques for estimating the economic For a user making decisions based on forecasts, each of
value of weather forecasts is given by Murphy (1994), the four outcomes in Table 1 has an expected cost, or
and a comprehensive review by Katz and Murphy loss, as given in Table 2. For convenience, it is normal
(1997). The method applied in this study is closely to measure all costs and losses relative to the user’s

308
14698080, 2002, 3, Downloaded from https://rmets.onlinelibrary.wiley.com/doi/10.1017/S1350482702003043 by Readcube (Labtiva Inc.), Wiley Online Library on [15/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Decision-making from probability forecasts
costs for a Correct Rejection, so the ‘Normal Loss’ N 4.1 Mean expense following forecasts
is set to zero. (Note: this assumption is not necessary,
and the method readily accounts for non-zero values of Given the information specified in Tables 1 and 2, and
N.) If the event occurs with no protective action being assuming the user takes protective action whenever the
taken, the user incurs a loss L. If the event is forecast to event is forecast, it can be expected that over a number
occur, the user is assumed to take protective action at of forecast occasions the user will experience a mean
cost C. In the simple cost-loss situation (Murphy expense Efx of
1977), this action is assumed to give full protection if
the event does occur, so the user incurs the cost C for Efx = hLm + mL + fC + rN (1)
both Hits and False Alarms (Lm=C). In reality, protec-
tion will often not be fully effective in preventing all The last term rN is normally zero, but this specifies the
losses, and the losses may be generalised by specifying generalisation to any definition of N.
a Mitigated Loss Lm for Hits, as in Table 2.

Table 2. Contingency table of generalised user costs 4.2 Climatological reference expense
and losses. Note: in the simple cost-loss situation
described by Murphy (1977), this is simplified such that Forecast value represents the saving the user gains by
Lm = C. using the forecasts, and therefore requires a baseline
reference expense for comparison with Efx. If no fore-
Event Forecast cast is available the user has two options: either always
protect or never protect. In Murphy’s (1977) simple
Yes – User No – No cost-loss situation, where Lm=C, these options will
Protects Protective incur mean expenses Ecl over many occasions of C or
Action ōL respectively, where ō is the climatological frequency
Yes Mitigated Loss L of occurrence of the event. The user’s best choice is
Loss Lm always to protect if C<ō L, or C/L<ō and never to pro-
Events tect otherwise. Assuming the user takes the best option,
Observed No Cost C Normal the mean climatological expense in the simple cost-loss
Loss N=0 situation is thus given by

Ecl = min(C,ōL) (2)


For a forecast to have value it is necessary that Lm<L. For the generalised user loss matrix given in Table 2,
In most circumstances it would be expected that the mean expense of the ‘always protect’ option is given
C≤ Lm<L, but it is possible that in some circum- by (1–ō)C + ōLm, and the mean expense of following
stances Lm<C. For example, protective action could climatology is given by
involve using an alternative process which works effec-
tively in the weather conditions specified by the event, Ecl = min((1–ō)C + ōLm , ōL) (3)
but does not work in the non-event conditions; in this
case the cost C of a False Alarm would be high com- Fully generalising this to allow for the possibility of
pared to Lm. N≠0 gives

The above examples assume that costs, losses and fore- Ecl = min((1–ō)C + ōLm ,(1 – ō)N + ōL) (4)
cast value are specified in monetary terms. They could,
instead, be calculated in terms of a user’s energy con- In this case the user’s best strategy is always to take
sumption – the concept is the same. One limitation of protective action if
the system comes where a forecast is used to protect
C–N
life, owing to the difficulty in placing a cost on lives a= < ō (5)
lost. L + C – Lm – N

where α is the generalised cost/loss ratio introduced by


4. Calculation of forecast value Richardson (2000). In some circumstances one of the
climatological options may not be viable for a user,
Forecast value will be defined first for a simple deter- since taking protective action may involve stopping
ministic forecast, and the generalisation to probability doing their normal economic activity (e.g. a fisherman’s
forecasts will be considered in more detail in the next protective action against strong winds may be to leave
section. Value is defined as the saving a user can make his boat in port). The user cannot do that all the time or
by using a forecast, so it is necessary to evaluate the he would go out of business. In this case the forecast
user’s mean expense, averaged over many occasions, value should be calculated using the viable climate
with and without the benefit of the forecast. option. This will be considered further in section 5.3.

309
14698080, 2002, 3, Downloaded from https://rmets.onlinelibrary.wiley.com/doi/10.1017/S1350482702003043 by Readcube (Labtiva Inc.), Wiley Online Library on [15/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Kenneth R. Mylne
Fig. Fig.
1a 1b

Figure 1. Examples of relative forecast value Vr, expressed as a percentage of the value of a perfect forecast, plotted against prob-
ability threshold pt for probability forecasts of an event. The values of equivalent deterministic forecasts are shown in the
columns at the right-hand side of each graph. The best climatological option, selected following equation (4), is labelled above
the graph. Details of the forecasts used are given in section 5. Forecast lead-times are: 48h (solid line), 96h (dotted), 144h (dot-
dot-dot-dash) and 192h (dashed). The three graphs are for the same forecasts, but for different user loss functions as given in
Table 3: (a) A; (b) B; and (c) C. (Note that values for deterministic forecasts at 144h and 192h in Figure 1c are missing because
they are off the graph below –40%.)

4.3 Definition of forecast value service though, a user will want to consider Cfx, and a
forecast provider may use equation (7) to place an
The value of the forecast in monetary terms, V, is the upper bound on Cfx.
saving the user can expect to make from following the
forecast, averaged over a number of occasions: For general assessments of forecast value, it is conve-
nient to scale V relative to the value of a perfect fore-
V = Ecl – Efx (6) cast, in a similar fashion to the standard definition of a
skill score (see Wilks 1995). With perfect forecasts
This basic definition of value is the most relevant for a m=f=0 and the user takes protective action only when
user, except that it does not account for the cost of buy- the event occurs. The mean expense is therefore Ep =
ing the forecast service, Cfx. The true value to the user ōLm (or if N≠0, then Ep = ōLm + (1 – ō)N.) The relative
is therefore economic value of a forecast is then defined as:

Vu = V – Cfx = Ecl – Efx – Cfx (6) Ecl – Efx


Vr = (8)
Ecl – Ep
However, Cfx depends on the pricing policy of the
provider and cannot be estimated in general terms, so as a maximum of 1 (or 100%) for a perfect forecast sys-
for the purposes of this paper this term will be ignored tem, and is zero for a climatology forecast.
and value defined as in equation (6). In considering any

310
14698080, 2002, 3, Downloaded from https://rmets.onlinelibrary.wiley.com/doi/10.1017/S1350482702003043 by Readcube (Labtiva Inc.), Wiley Online Library on [15/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Decision-making from probability forecasts

Fig. Table 3. User costs and losses in arbitrary monetary


1c units for the three examples of forecast value plots
shown in Figure 1. (For the meaning of the column
headings, see Tables 1 and 2.)

User N Lm L C
(CR) (H) (M) (FA)
A 0 1 5 1
B 0 1 2 1
C 0 2 10 1

produce perfectly reliable forecasts, so it would be use-


ful to find a way of identifying the ideal probability
threshold for real, imperfect probability forecasts.

Forecast value is defined above for deterministic fore-


casts. To make decisions from probability forecasts, the
user takes protective action (‘Forecast=Yes’ in Table 1)
when the probability p of the event exceeds a chosen
threshold pt. The value of the forecast therefore
depends on the choice of pt. To specify completely the
value of a probability forecast system, value must be
calculated for a range of probability thresholds and
plotted as a function of pt as shown in Figure 1. (This
use of probability thresholds is identical to that used in
ROC verification, for which hit rates (HR) and false
alarm rates (FAR) are calculated for various probability
thresholds from the contingency table in Table 1. The
appendix describes how forecast value may be evalu-
ated immediately for any forecast system for which
ROC verification is available.)

While Vr has a maximum of 1, there is no lower limit, Three examples of forecast value plots are shown in
and from equations (1) and (8) it is clear that negative Figure 1. All are based on verification of the same fore-
values are likely when m or f, or their corresponding casts, but the user costs and losses are different, as
user losses Lm or C, are large. Negative value simply shown in Table 3. The verification is based on daily
indicates that the forecast system does not have suffi- forecasts of the event ‘10 m wind speed of Beaufort
cient skill to provide value to this particular user. In Force 5 or more’ for 41 sites in the UK, over two win-
this case the user’s best strategy is to follow the best cli- ter seasons (December, January, February, 1998/99 and
matological option. 1999/2000). These forecasts were generated and veri-
fied by the Met Office’s PREVIN system, which is
described in detail by Legg et al. (2001). Probability
5. Decision-making and value for probability forecasts were taken from the 51-member ECMWF
forecasts EPS; for comparison, equivalent deterministic forecasts
from the ECMWF high resolution (TL319) model are
Many forecast users are used to making decisions from also included. Results are shown for forecast lead-times
deterministic forecasts, which give them a simple yes/no of 48, 96, 144 and 192 hours. The value of the probabil-
answer on whether to take protective action against the ity forecasts is calculated and plotted at probability
weather, and are uncertain how to respond to a proba- thresholds of 0, 10, 20, 30, …, 90, 100%.
bility forecast. Given the cost-loss information in Table
2, the probability level at which protection should be
taken can be easily related to the relative cost and bene- 5.1 Value of probability forecasts
fit of protection. In the simple cost-loss situation (Lm =
C), the user’s best strategy should be to protect when- From Figure 1a it can be seen that the value of a prob-
ever the probability of the event exceeds C/L; in the ability forecast is strongly dependent on pt. (The form
more general cost-loss situation, the threshold is α, as of this function is identical for V and Vr, since the scal-
defined in equation (5). However, this assumes that the ing of V by Ecl-Ep in equation (8) is independent of pt.)
forecasts perfectly represent the true probability of the The user’s requirement is to get maximum value from
event. In reality probability forecast systems do not the system, and therefore the practical value of the

311
14698080, 2002, 3, Downloaded from https://rmets.onlinelibrary.wiley.com/doi/10.1017/S1350482702003043 by Readcube (Labtiva Inc.), Wiley Online Library on [15/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Kenneth R. Mylne
probability forecasts is given by the maximum of the
curve at pt= pmax. By using pmax as a decision threshold,
the user can optimise future decision-making from the
forecast system (provided only that past performance is
representative of future performance); for an imperfect
system pmax may be different from the probability
threshold expected from α, but will give the user the
maximum benefit from the available forecast informa-
tion. The form of the graphs in Figure 1 allows easy
identification of pmax for a particular user, and also
allows direct comparison of the value of the probabil-
ity forecast Vr(pmax) with the value of an equivalent
deterministic forecast.

For User A (Figure 1a), α=C/L=0.2 so pmax should be


20% for perfectly reliable forecasts. In practice, the
value is a maximum for probability thresholds of 20%
for lead-times of 144h and 192h, but 30% at 48h and
96h. The value falls negative for large thresholds.

For User B (Figure 1b), α=C/L=0.5 but maximum


value is obtained with pt of 70–80%. The only differ- Figure 2. Reliability diagram for the probability forecasts
used to generate the forecast value graphs of Figure 1. Line
ence between users A and B is that L for a missed event
types indicate the forecast lead-times as in Figure 1.
is greater for A. User A therefore benefits more by
lowering the miss rate, which is achieved by reducing pt at 192h, maximum value is obtained with pmax =10%,
and thus increasing the number of occasions when pro- indicating that, with a 51-member ensemble, the user
tective action is taken. For User B, protective action is should take protective action when only 5 ensemble
relatively expensive so protection is only advisable members predict the event to occur. For 96h and 144h
when the probability of the event is quite high. forecasts, the value is still increasing rapidly as the
threshold decreases to 10%, and it appears likely that
A reliability diagram for the probability forecasts used even better value could be obtained using a threshold of
in Figure 1 is shown in Figure 2. Forecasts are cate- 2 or 3 ensemble members. For the 192h forecasts the
gorised by the forecast probability, and the frequency value is negative even at pt=10%. Again, it is possible
of occurrence of the event for each category plotted that a small positive benefit could be obtained from the
against the forecast probability. For a perfectly reliable forecasts using a lower threshold. In this study forecast
system the points will lie along the diagonal line, but it verification was only available at 10% probability
can be seen that for these forecasts the slope is rather thresholds, so it is not possible to confirm this. For
less than 45°, so that for high forecast probabilities the practical applications where a user has a low cost-loss
event occurs less frequently than predicted. It is clearly ratio α, it is important to carry out the verification for
seen that observed frequencies of 50% occur for fore- all possible low probability thresholds down to a single
cast probabilities of 70–80%, in agreement with pmax ensemble member, in order to ensure optimum use of
for User B. Similarly for User A, for lead-times of 48h the ensemble information. This is likely to be particu-
and 96h, the observed frequency is close to 20% when larly important when considering the use of ensembles
the forecast probability is 30%. A representative relia- for predicting severe weather events, when the losses
bility diagram such as that in Figure 2 can be used to from not protecting are often high.
calibrate forecast probabilities and produce near-per-
fectly reliable forecasts. Zhu et al. (2001) recommend It must be noted that for some users the forecast value
the use of calibrated probabilities so that the ideal deci- may always be negative, even though the forecast has
sion threshold can be deduced directly from C/L (or skill, because the costs of misses and false alarms are
α). The method described in this paper avoids the need too high relative to the savings made from a correct
for separate calibration, and allows the user (working forecast. In this case the user’s best strategy is to either
with the forecast provider) to obtain the maximum always or never protect, whichever is better, rather
benefit from any probability forecast system. In effect, than following the forecasts.
calibration is built into the identification of pmax.

The example of User C shown in Figure 1c is similar to 5.2 Comparison of probability and deterministic
User A, except that the false alarm cost is half that of forecasts
User A. This example demonstrates the flexibility of
the system in allowing Lm ≠ C. The expected decision Also shown in Figure 1 are the values of the determin-
threshold is α=0.111, and Figure 1c shows that, except istic forecasts from the higher resolution TL319
312
14698080, 2002, 3, Downloaded from https://rmets.onlinelibrary.wiley.com/doi/10.1017/S1350482702003043 by Readcube (Labtiva Inc.), Wiley Online Library on [15/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Decision-making from probability forecasts
ECMWF model, calculated in exactly the same way, 6. Practical application
but with only a single yes/no boundary instead of a
range of probability thresholds. This provides a direct Practical application of this technique for the use of
comparison, in terms which are relevant to the user, of probability forecasts as decision-making tools will
the benefits of deterministic and probabilistic forecasts. require close collaboration between forecast providers
In the case of the 48-hour forecasts for User A the and users. The user is required to define the event to be
deterministic value is almost identical to the maximum forecast, and also to estimate the losses expected in each
value of the probability forecast; in every other case of the contingencies. The forecast provider must pro-
shown the probability forecast has more user value vide verification of the forecast system’s ability to pre-
than the equivalent deterministic forecast, despite the dict the defined event. Given these data, forecast value
higher resolution of the deterministic model. In many functions can be plotted; pmax can then be identified,
cases the probability forecast has positive value when and the user takes protective action whenever the fore-
the deterministic forecast has negative value. This is cast probability exceeds this threshold.
largely due to the fact that since the deterministic fore-
cast can generate probabilities of only 0 or 100%, it Although this process is simple in principle, it is less
cannot adapt to the requirements of different users. By straightforward in practice. Many decision-making
contrast, the probability forecast can use a range of processes are more complex than a simple
probabilities, and can be interpreted appropriately for ‘protect’/‘don’t protect’ choice. Estimation of users’
each user as explained above (see also Zhu et al. 2001). losses is not easy, and verification data may not be
immediately available for the defined event. It has been
For some forecast events and users the probability assumed here that the user will want to base decisions
forecast can have more value even at 12 or 24 hours, on maximising savings over a large number of occa-
although in general the benefits are greater at longer sions. Some users may place their emphasis on reducing
lead-times. For longer-range forecasts there is inher- their exposure to large losses; however, it should be
ently more uncertainty, and a deterministic forecast is possible to reflect this in suitable choices of the costs
likely to be wrong more frequently, whereas a proba-
bility forecast can account for the uncertainty by using
mid-range probabilities. Occasionally it is found that
the deterministic forecast is more valuable, even at
quite long lead-times. This occurs when the behaviour
of the deterministic model for the defined event is well-
tuned to the particular user’s losses. In any case, the
method allows the best strategy for any user to be iden-
tified, and in most cases this will be from the probabil-
ity forecasts.

5.3 Effect of the climatological reference expense


From Figure 1 it can be seen that in cases A and C the
value of the probability forecasts converges to zero at pt
=0 and in case B at pt =1. The points at pt =0 and 1 cor-
respond to the event always and never being predicted
respectively, so that V=0 at one of these points because
Efx=Ecl.

In general, the correct climatological reference option


to use will be the cheapest, as described in section 4.2.
This minimises the forecast value and avoids any risk of
over-estimating it. However, in some real situations the
user may not have that option, as described at the end
of section 4.2. For such users it would be appropriate to
choose a fixed climatological baseline for the forecast
value, rather than the minimum given by equation (4).
In Figure 3, the example of Figure 1a is shown assum-
ing that the user does not have an option to ‘always
protect’. In this case it is found that the relative value of
the forecasts is now much greater. Since it is only Ecl
which has changed in equation (8), and Ecl>Ep, the Figure 3. As Figure 1a, but with the climatological reference
absolute value V is also larger. expense restricted to the ‘never protect’ option.

313
14698080, 2002, 3, Downloaded from https://rmets.onlinelibrary.wiley.com/doi/10.1017/S1350482702003043 by Readcube (Labtiva Inc.), Wiley Online Library on [15/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Kenneth R. Mylne
and losses, with the result that the system will identify Carter, G. M., Dallavalle, J. P. & Glahn, H. R. (1989)
a low protection threshold pmax. Despite these difficul- Statistical forecasts based on the National Meteorological
ties, it is hoped that this method will help to develop Center’s numerical weather prediction system. Wea. and
better understanding of how probability forecasts can Forecasting 4: 401–412.
be used in decision-making, and lead to further work Glahn, H. R. & Lowry, D. A. (1972) The use of model out-
put statistics (MOS) in objective weather forecasting. J.
on less idealised decision-making scenarios. Benefits
Appl. Meteorol. 11: 1203–1211.
are likely to be greatest for users with large weather- Graham, R. J., Evans, A. D. L., Mylne, K. R., Harrison, M. S.
sensitivity and large potential losses, for whom the J. & Robertson, K. B. (2000) An assessment of seasonal
investment in initial analysis is justified by the returns. predictability using atmospheric general circulation mod-
els. Q. J. R. Meteorol. Soc. 126: 2211–2240.
Forecast providers need to be aware that analysis of Hamill, T. M. & Colucci, S. J. (1997) Verification of Eta-RSM
some customers’ requirements may reveal that it is not short-range ensemble forecasts. Mon. Wea. Rev. 125:
possible to provide a forecast with value, in which case 1312–1327.
the customer is best advised not to purchase the forecast! Hamill, T. M. & Colucci, S. J. (1998) Evaluation of Eta-RSM
Conversely, Graham et al. (2000) demonstrate that the ensemble probabilistic precipitation forecasts, Mon. Wea.
technique may be used to gain useful information from Rev. 126: 711–724.
Houtekamer, P. L., Lefaivre, L, Derome, J., Ritchie, H. &
seasonal forecasts for which forecast skill measured by
Mitchell, H. L. (1996) A system simulation approach to
conventional verification techniques, such as ROC, is ensemble prediction. Mon. Wea. Rev. 124: 1225–1242.
relatively low compared to medium-range forecasts. Katz, R. W. & Murphy, A. H. (eds.) (1997) Economic Value
of Weather and Climate Forecasts. Cambridge University
Press, Cambridge.
7. Conclusions Legg, T. P., Mylne, K. R. & Woolcock, C. (2001) The use of
medium-range ensembles at the Met Office. I: PREVIN. A
A technique for estimating forecast value has been system for the production of probabilistic forecast infor-
mation from the ECMWF EPS, to appear in Meteorol.
described which allows direct comparison of the value
Appl.
of probability and deterministic forecasts in terms Lorenz, E. N. (1963) Deterministic nonperiodic flow. J.
which are relevant to forecast users. Probability fore- Atmos. Sci. 20: 130–141.
casts based on the ECMWF EPS are usually found to Lorenz, E. N. (1993) The Essence of Chaos. UCL Press Ltd,
have greater value than equivalent deterministic fore- London, 227pp.
casts from a higher resolution model, particularly for Molteni, F., Buizza, R., Palmer, T. N. & Petroliagis, T. (1996)
forecasts with lead-times greater than about 48 hours. The ECMWF ensemble prediction system: methodology
and validation. Q. J. R. Meteorol. Soc. 122: 73–119.
The technique may be used to apply probability fore- Murphy, A. H. (1977) The value of climatological, categorical
casts as decision-making tools by identifying a proba- and probabilistic forecasts in the cost-loss situation. Mon.
bility threshold which maximises the forecast value. Wea. Rev. 105: 803–816.
Murphy, A. H. (1994) Assessing the economic value of
The forecast user then takes protective action against
weather forecasts: an overview of methods, results and
the event whenever the forecast probability exceeds issues. Meteorol. Appl. 1: 69–74.
this threshold. Richardson, D. (2000) Skill and relative economic value of the
ECMWF ensemble prediction system Q. J. R. Meteorol.
Soc. 126: 649–667 (January, Part B).
Acknowledgements Stanski, H. R., Wilson, L. J. & Burrows, W. R. (1989) Survey
of Common Verification Methods in Meteorology. WMO
The author would like to thank Tim Legg and Caroline WWW Tech Report No 8, WMO TD No. 358.
Woolcock for providing the ensemble verification data Thornes, J. E. & Stephenson, D. B. (2001) How to judge the
used in this study, and Mike Harrison for stimulating quality and value of weather forecast products. Meteorol.
the ideas behind the project. The author also thanks the Appl. 8: 307–314.
Toth, Z. & Kalnay, E. (1993) Ensemble forecasting at the
two anonymous referees for their constructive com-
NMC: the generation of perturbations. Bull. Amer.
ments on the first draft of this paper. Meteor. Soc. 74: 2317–2330.
Toth, Z., Kalnay, E., Tracton, S. M., Wobus, R. & Irwin, J.
(1997) A synoptic evaluation of the NCEP ensemble, Wea.
References and Forecasting 12: 140–153.
Wilks, D. S. (1995) Statistical Methods in the Atmospheric
Brier, G. W. (1950) Verification of forecasts expressed in Sciences: An Introduction, International Geophysics Series,
terms of probability. Mon. Wea. Rev. 78: 1–3. Vol. 59, Academic Press.
Buizza, R. & Palmer, T. N. (1998) Impact of ensemble size on Zhu, Y., Toth, Z, Wobus, R., Richardson, D. & Mylne, K.
ensemble prediction. Mon. Wea. Rev. 126: 2503–2518. (2001) On the economic value of ensemble-based weather
Buizza, R., Miller, M. & Palmer, T. N. (1999) Stochastic rep- forecasts, forthcoming in the Bulletin of the American
resentation of model uncertainties in the ECMWF EPS. Q. Meteorological Society.
J. R. Meteorol. Soc. 125: 2887–2908.

314
14698080, 2002, 3, Downloaded from https://rmets.onlinelibrary.wiley.com/doi/10.1017/S1350482702003043 by Readcube (Labtiva Inc.), Wiley Online Library on [15/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Decision-making from probability forecasts
Appendix. The link between forecast value and
H+M
ROC verification ō = (A4)
H+M+F+R
The contingency table in Table 1 is the same as that
used in ROC verification (see Stanski et al. 1989). ROC Hence, combining equations (A1), (A3) and (A4):
expresses the forecast performance in terms of a hit rate
(HR) and false alarm rate (FAR) defined as: h = ōHR (A5)
H
HR = (A1) Similarly:
H+M
m = ō(1– HR) (A6)
F
FAR= (A2) f = (1 – ō)FAR (A7)
F+R
r = (1 – ō)(1 – FAR) (A8)
where H,M,F and R are as defined in Table 1.
In ROC verification of probability forecasts HR and
To evaluate Efx from equation (1) we require values of FAR are evaluated for a range of probability thresholds
h, m, f and r. By definition, the frequency of hits h is in exactly the same way as described in this paper for
forecast value. Thus equation (1), and hence the fore-
H
h= (A3) cast value, are easily evaluated for any forecast for
H+M+F+R which ROC verification data are available and the
climatological frequency of the event, ō, is known.
and the climatological frequency of the event is

315

You might also like