calcscore: Calculate Scores Under A Specific Rule

In scoring: Proper Scoring Rules

Description

Given parameters of a scoring rule family, calculate scores for probabilistic forecasts and associated outcomes.

Usage

## S3 method for class 'formula'
calcscore(object, fam="pow", param, data, bounds=NULL,
          reverse=FALSE, ordered=FALSE, ...)

## Default S3 method:
calcscore(object, outcome, fam="pow",
          param=c(2,rep(1/max(2,NCOL(forecast)),max(2,NCOL(forecast)))),
          bounds=NULL, reverse=FALSE, ordered=FALSE, ...)

Arguments

object	an object of class "formula", of the form outcome ~ forecast (see details). Alternatively, a matrix of forecasts, with observations in rows and forecast alternatives in columns. For two-alternative forecasts, this can be a vector reflecting forecasts for one alternative.
outcome	a vector of outcomes (used if object is a matrix). For each row of the forecast matrix, outcome should contain an entry reflecting the column number associated with the event that occurred.
fam	scoring rule family. pow (default) is the power family, beta is the beta family, sph is the pseudospherical family.
param	for family beta, a numeric vector of length 2 containing the scoring rule family parameters. For other families, a numeric vector containing first the family parameter gamma and optionally NCOL(forecast) baseline parameters (see details). Alternatively, a matrix may be supplied containing unique family parameters for each forecast row.
data	an optional data frame or list containing the variables in the formula. If not found in data, the variables are taken from the environment from which calcscore is called.
bounds	a vector of length 2 corresponding to the desired minimum value and maximum value of the scoring rule, respectively. Entries of NA imply that the minimum and/or maximum bound will not be modified from the natural, family-implied bounds.
reverse	if FALSE (default), smaller scores imply better forecasts. If TRUE, larger scores imply better forecasts.
ordered	if FALSE (default), forecast alternatives have no ordering. If TRUE, forecast alternatives have the ordering implied by forecast. The resulting scoring rule is sensitive to this ordering (see details).
...	Additional arguments.

Details

The formula is of the form outcome ~ forecast, where forecast describes the column(s) containing forecasts associated with the possible outcomes. Multiple columns are separated by +. outcome is always a vector describing the outcome associated with each forecast. It should be coded 1, 2, ..., reflecting the column associated with the outcome (see examples).

For events with only two alternatives, one can take a shortcut and supply only forecasts associated with a single outcome (if baseline parameters are specified for families pow and sph, the parameter for only that outcome should also be supplied). In this case, the outcome vector should contain zeros and ones, where ‘one’ means that the forecasted alternative occurred.

If ordered=TRUE, an "ordered" scoring rule is obtained using the strategy proposed by Jose, Nau, & Winkler (2009). These ordered rules are only useful when the number of forecasted alternatives is greater than two (i.e., when one uses family pow or sph).

If baseline parameters are not supplied for families pow or sph, then the parameters are taken to be equal across all alternatives (though the natural scaling of the scoring rule differs depending on whether or not one explicitly supplies equal baseline parameters).

If desired, a unique scoring rule can be applied to each row of the forecast matrix: the param argument can be supplied as a matrix.

When the bounds argument is supplied, the code attempts to scale the scores so that the maximum score is bounds[2] and the minimum score is bounds[1]. This scaling cannot be accomplished when the scoring rule allows scores of infinity (the log score is the most common case here). If reverse=TRUE, the bounds are applied after the reversal (so that the supplied lower bound reflects the worst score and upper bound reflects the best score).

Value

calcscore returns a numeric vector that has length equal to length(outcome), containing scores under the selected scoring rule.

Note

The beta family was originally proposed by Buja et al.\ (2005); the power and pseudospherical families with baseline are described by Jose et al.\ (2009). A discussion of choosing specific rules from these families is provided by Merkle and Steyvers (2013).

Some notable special cases of these families are:

Beta family: Log score when parameters are (0,0); Brier score when parameters are (1,1).

Power family with baseline parameters all equal (to 1/(number of alternatives)): The family approaches the log score as gamma goes to 1 (but the family is undefined for gamma=1). The Brier score is obtained for gamma=2.

Pseudospherical family with baseline parameters all equal: The family approaches the log score as gamma goes to 1 (but the family is undefined for gamma=1). The spherical score is obtained for gamma=2.

Author(s)

Ed Merkle

References

Buja, A., Stuetzle, W., & Shen, Y. (2005). Loss functions for binary class probability estimation and classification: Structure and applications. (Obtained from http://stat.wharton.upenn.edu/~buja/PAPERS/)

Jose, V. R. R., Nau, R. F., & Winkler, R. L. (2008). Scoring rules, generalized entropy, and utility maximization. Operations Research, 56, 1146–1157.

Jose, V. R. R., Nau, R. F., & Winkler, R. L. (2009). Sensitivity to distance and baseline distributions in forecast evaluation. Management Science, 55, 582–590.

Merkle, E. C. & Steyvers, M. (in press). Choosing a strictly proper scoring rule. Decision Analysis.

See Also

plotscore

Examples

## Brier scores for two alternatives, with bounds of 0 and 1
data("WorldEvents")
scores <- calcscore(answer ~ forecast, fam="beta",
                    param=c(1,1), data=WorldEvents,
                    bounds=c(0,1))

## Calculate Brier scores manually
scores.man <- with(WorldEvents, (forecast - answer)^2)

## Comparison
all.equal(scores, scores.man)

## Average Brier score for each forecaster
with(WorldEvents, tapply(scores, forecaster, mean))

## Brier scores for 3 alternatives, with bounds of 0 and 1
data("WeatherProbs")
scores2 <- calcscore(tcat ~ tblw + tnrm + tabv, fam="pow",
                     param=2, data=WeatherProbs,
                     bounds=c(0,1))

## Spherical scores for 3 alternatives, reversed so 0 is worst and
## 1 is best
scores3 <- calcscore(tcat ~ tblw + tnrm + tabv, fam="sph",
                     param=2, data=WeatherProbs,
                     bounds=c(0,1), reverse=TRUE)

## Replicate Jose, Nau, & Winkler, 2009, Figure 1
r2 <- seq(0, .6, .05)
r <- cbind(.4, r2, .6 - r2)
j <- rep(1, length(r2))

## Panel 1
quad <- calcscore(j ~ r, fam="pow", param=2, bounds=c(-1,1), reverse=TRUE)
quadbase <- calcscore(j ~ r, fam="pow", param=c(2,.3,.6,.1), reverse=TRUE)
rankquad <- calcscore(j ~ r, fam="pow", param=2, ordered=TRUE, reverse=TRUE)
rankquadbase <- calcscore(j ~ r, fam="pow", param=c(2,.3,.6,.1), ordered=TRUE,
                          reverse=TRUE)
plot(r2, quad, ylim=c(-2,1), type="l", ylab="Quadratic scores")
lines(r2, quadbase, lty=2)
lines(r2, rankquad, type="o", pch=22)
lines(r2, rankquadbase, type="o", pch=2)

## Panel 2
sph <- calcscore(j ~ r, fam="sph", param=2, reverse=TRUE, bounds=c(-1.75,1))
sphbase <- calcscore(j ~ r, fam="sph", param=c(2,.3,.6,.1), reverse=TRUE)
ranksph <- calcscore(j ~ r, fam="sph", param=2, ordered=TRUE, reverse=TRUE)
ranksphbase <- calcscore(j ~ r, fam="sph", param=c(2,.3,.6,.1), ordered=TRUE,
                         reverse=TRUE)
plot(r2, sph, ylim=c(-1,.6), type="l", ylab="Spherical scores")
lines(r2, sphbase, lty=2)
lines(r2, ranksph, type="o", pch=22)
lines(r2, ranksphbase, type="o", pch=2)

## Panel 3
lg <- calcscore(j ~ r, fam="pow", param=1.001, reverse=TRUE)
lgbase <- calcscore(j ~ r, fam="pow", param=c(1.001,.3,.6,.1), reverse=TRUE)
ranklg <- calcscore(j ~ r, fam="pow", param=1.001, ordered=TRUE, reverse=TRUE)
ranklgbase <- calcscore(j ~ r, fam="pow", param=c(1.001,.3,.6,.1),
                        ordered=TRUE, reverse=TRUE)
plot(r2, lg, ylim=c(-2,1), type="l", ylab="Log scores")
lines(r2, lgbase, lty=2)
lines(r2, ranklg, type="o", pch=22)
lines(r2, ranklgbase, type="o", pch=2)

Example output

[1] "Mean relative difference: 9.537645e-07"
         1          2          3          4          5          6          7 
0.42143333 0.02136665 0.27086667 0.33333333 0.24000000 0.21496667 0.28623333 
Warning message:
In calcscore.default(object = list(tblw = c(0.4341, 0.4243, 0.4443,  :
  Forecasts in some rows do not sum to 1; they were scaled to sum to 1.
Warning message:
In calcscore.default(object = list(tblw = c(0.4341, 0.4243, 0.4443,  :
  Forecasts in some rows do not sum to 1; they were scaled to sum to 1.

scoring documentation built on May 2, 2019, 4:53 p.m.