Shadow and validation · February initialization · FMAM and its months
Shadow forecast
Not a published forecast. For each new cycle of this initialization, the shadow method runs through the same probability pipeline as the published forecast (all reference years; blend weight from leave-one-year-out), next to it, and is scored once CHIRPS for the target is available. Each scored season is independent evidence that enters the suggestion.
Feb 2026: Square-root-space Mean-Variance Scaling
The published method run alongside reproduces the published forecast: largest probability difference 1.9e-8, blend weight difference 5.7e-9.
| Method | Region | RPSS | BSS below | BSS above | Amount CRPSS | Favoured category observed | λ |
|---|---|---|---|---|---|---|---|
| Mean-Variance Scaling (published) | All Ethiopia | -0.012 | -0.074 | +0.062 | +0.015 | 40% | 0.50 |
| Mean-Variance Scaling (published) | FMAM rainfall contribution domain | +0.002 | -0.051 | +0.063 | +0.020 | 42% | 0.50 |
| Square-root-space Mean-Variance Scaling (shadow) | All Ethiopia | +0.003 | -0.059 | +0.077 | +0.028 | 40% | 0.35 |
| Square-root-space Mean-Variance Scaling (shadow) | FMAM rainfall contribution domain | +0.037 | -0.025 | +0.109 | +0.042 | 44% | 0.35 |
Independent validation seasons
Every method scored on the seasons after the benchmark (one season each; the suggestions need at least three).
| Method | Feb | Mar | Apr | May | FMAM |
|---|---|---|---|---|---|
| Mean-Variance Scaling (operational) | -0.012 | +0.057 | -0.134 | +0.002 | -0.013 |
| CDF-transform | -0.012 | +0.127 | -0.134 | -0.015 | +0.000 |
| Delta Change | +0.005 | +0.041 | -0.114 | -0.034 | -0.020 |
| Detrended Quantile Mapping | +0.008 | +0.085 | -0.128 | -0.027 | +0.006 |
| Empirical Quantile Mapping | -0.002 | +0.081 | -0.132 | -0.016 | -0.013 |
| Gamma Quantile Mapping | -0.071 | +0.092 | -0.156 | -0.007 | -0.004 |
| Local Intensity Scaling | -0.004 | +0.001 | -0.151 | -0.002 | -0.016 |
| Linear Scaling | +0.004 | +0.007 | -0.159 | -0.002 | -0.016 |
| MVS with all-zero guard (variant) | -0.012 | +0.057 | -0.134 | +0.002 | -0.013 |
| Power Transformation | -0.013 | +0.116 | -0.168 | -0.009 | -0.003 |
| Quantile Delta Mapping | +0.018 | +0.105 | -0.135 | -0.025 | +0.011 |
| Scaled Distribution Mapping | -0.001 | +0.004 | -0.116 | -0.020 | +0.006 |
| Square-root-space Mean-Variance Scaling | +0.003 | +0.100 | -0.170 | -0.005 | -0.008 |
| Variance Scaling (unbounded) | +0.005 | +0.059 | -0.131 | +0.003 | -0.014 |
| Method | Feb | Mar | Apr | May | FMAM |
|---|---|---|---|---|---|
| Mean-Variance Scaling (operational) | +0.015 | +0.067 | -0.486 | -0.032 | -0.091 |
| CDF-transform | -0.343 | +0.133 | -0.633 | -0.075 | -0.090 |
| Delta Change | -0.165 | +0.077 | -0.479 | -0.061 | -0.099 |
| Detrended Quantile Mapping | -0.211 | +0.102 | -0.591 | -0.085 | -0.080 |
| Empirical Quantile Mapping | +0.080 | +0.082 | -0.507 | -0.054 | -0.099 |
| Gamma Quantile Mapping | -0.008 | +0.094 | -0.490 | -0.057 | -0.080 |
| Local Intensity Scaling | -0.282 | +0.092 | -0.574 | -0.091 | -0.106 |
| Linear Scaling | -0.351 | +0.093 | -0.579 | -0.093 | -0.107 |
| MVS with all-zero guard (variant) | +0.015 | +0.067 | -0.486 | -0.032 | -0.091 |
| Power Transformation | -0.060 | +0.110 | -0.500 | -0.048 | -0.088 |
| Quantile Delta Mapping | -0.115 | +0.129 | -0.661 | -0.088 | -0.097 |
| Scaled Distribution Mapping | -0.394 | +0.018 | -0.581 | -0.064 | -0.107 |
| Square-root-space Mean-Variance Scaling | +0.028 | +0.091 | -0.495 | -0.051 | -0.085 |
| Variance Scaling (unbounded) | +0.024 | +0.067 | -0.489 | -0.031 | -0.094 |