Scoring epidemiological forecasts on transformed scales
Fig 3
Illustration of the effect of the log-transformation of the ranking for a single forecast.
Shown are CRPS (or WIS, respectively) values as a function of the observed value for two forecasters. Model A issues a geometric distribution (a negative binomial distribution with size parameter θ = 1) with mean μ = 10 and variance σ2 = μ + μ2 = 110), while Model B issues a Poisson distribution with mean and variance equal to 10. Zeroes in this illustrative example were handled by adding one before applying the natural logarithm.