Model v8.0: recalibrated on an out-of-sample test, and the counts that follow from it
The projection model was re-fitted and tested by fitting cohort change ratios on Census 2001 to 2011, forecasting 2021, and scoring against the actual Census 2021, so the fitting window never touches the target. On the White British share it scores mean absolute error 1.56pp with a bias of +0.05pp across 285 areas. The settings it replaces scored 2.82pp with a bias of -2.13pp on the same test, meaning the previous model projected change too fast rather than too slow. Shrinkage toward the national ratio is K=25 and the growth ceiling is 1.65 per decade, both chosen on that test. Counts move with it: 86 local authorities are projected below a 50% White British share by 2051 rather than 109, 59 of them currently majority, and Birmingham's 2051 share is 14.9% rather than 10.7%. The backcast comparison with NEWETHPOP is withdrawn; the defensible comparison is out-of-sample against out-of-sample, this model at 1.56pp across 285 areas against NEWETHPOP at 3.95pp across 296. Only a one-decade horizon has been tested.