How accurate are the estimates?
We hide reported salaries, estimate them with the model, and compare. Every test leaves the tested club out of the fit.
Main results
| Test | Players | Median abs. error | Within ±25% | Within ±50% |
|---|---|---|---|---|
| Only the club payroll known (leave one club-season out) | 222 | 28.8% | 44.6% | 78.4% |
| Payroll + top 25% of earners known (residual) | 166 | 21.4% | 54.2% | 81.3% |
| Payroll + top 50% of earners known (residual) | 111 | 22.8% | 55.0% | 85.6% |
| Other clubs, partial information (press payroll, no other salaries) | 87 | 38.6% | 28.7% | 64.4% |
| Salary bands only (role × budget tier) | 309 | 39.0% | 33.0% | 62.5% |
Median absolute error: half of the estimates are closer than this to the reported salary. On the 87 players at other clubs the model underestimates by 17.4% on the median: mostly stars at clubs that publish nothing.
Which inputs matter
Payroll only, same 222 players, adding inputs step by step:
| Model | Median abs. error | Within ±25% | Within ±50% | Median bias |
|---|---|---|---|---|
| F0 equal split | 41.4% | 28.4% | 58.1% | 20.4% |
| F1 minutes only | 31.4% | 39.6% | 71.6% | 8.3% |
| F2 + current PIR | 31.5% | 41.9% | 71.2% | 7.7% |
| F3 + prior season | 29.6% | 42.3% | 74.3% | 5.8% |
| F4 + age | 28.8% | 44.6% | 76.6% | 9.4% |
| F5 + nationality | 28.8% | 44.6% | 78.4% | 8.0% |
| F0b minutes share | 44.7% | 28.4% | 54.1% | 1.6% |
With a payroll that is itself off by up to ±15%, the median error is 28.3%. Using the BasketNews payroll instead of the sum of reported salaries: 33.1% (164 players).
Residual sharing, more cases
| Salaries known | Players tested | Median abs. error | Within ±25% | Within ±50% |
|---|---|---|---|---|
| top 25% known | 166 | 21.4% | 54.2% | 81.3% |
| top 50% known | 111 | 22.8% | 55.0% | 85.6% |
| top 75% known | 56 | 14.1% | 73.2% | 91.1% |
| random 25% known | 3320 | 27.5% | 45.6% | 78.1% |
| random 50% known | 2220 | 26.3% | 47.9% | 78.8% |
| random 75% known | 1120 | 22.6% | 53.7% | 83.8% |
"Random" rows repeat the test 20 times with a random set of known salaries, so the player count is larger.
Bands against the model
| Method | Players | Median abs. error | Within ±25% | Within ±50% |
|---|---|---|---|---|
| Band midpoint | 364 | 41.3% | 29.4% | 60.4% |
| Band, scaled to payroll (clipped) | 309 | 39.0% | 33.0% | 62.5% |
| Band, scaled to payroll | 309 | 37.5% | 34.0% | 63.4% |
| Share model | 309 | 30.9% | 40.1% | 74.4% |
| Blend of both | 309 | 29.3% | 41.7% | 71.5% |
The site uses the share model (with the residual step). Bands appear only as context.