The replay
The model run on every Alzheimer's readout that already happened, published once and hashed.
Method
The corpus is every Alzheimer's Phase 2 or 3 interventional drug trial on ClinicalTrials.gov whose first primary outcome has a posted analysis with a p-value: 81 trials, 20 of which hit. Each is priced as of its primary completion date. The sponsor's record counts only readouts posted before that date and PubMed counts the 24 months before it; the registry's design fields do not change at readout.
Score
| Brier, placebo-1 | 0.1895 |
| Brier, always the base rate | 0.1859 |
| Brier, always 50% | 0.2500 |
| sha256 | 227ee012a5f02ade9bff8db65b5bb64fd3cfba9142e9e93d175b8e176a62d715 |
The model is well ahead of a coin flip and level with the base rate. It is a back-test on data that existed before the model did, so the live record is the test that counts. Every row.