The forecast agent
A written scoring model, run every night at 02:00 UTC. No language model is called.
Inputs
Five public inputs, each a link in the forecast JSON: the registry record (phase, enrolment, sponsor class, endpoint), the target's Open Targets association with Alzheimer's (MONDO_0004975), the sponsor's earlier posted readouts, the class base rate from the pipeline reviews, and PubMed papers on the drug in the last 24 months.
The formula
log-odds = logit(prior) + 1.10 (G - 0.30) + 0.30 clamp(ln(n/300), -1.5, 1.5)
+ 0.20 industry + 1.20 ((hits+1)/(readouts+4) - 0.25) + 0.12 ln(1 + papers/20)
p = clamp(logistic(log-odds), 0.03, 0.92)| Term | Definition |
|---|---|
| Class prior | logit(prior[endpoint][phase]) |
| Genetic support | 1.10 × (G − 0.30), G = highest Open Targets association of the drug's targets with Alzheimer's |
| Size | 0.30 × clamp(ln(n / 300), −1.5, 1.5), n = enrolment |
| Sponsor class | 0.20 if the lead sponsor is industry |
| Sponsor record | 1.20 × ((hits + 1) / (readouts + 4) − 0.25), the sponsor's posted readouts before tonight |
| Literature | 0.12 × ln(1 + papers / 20), PubMed papers on the drug and Alzheimer's in the 24 months before |
The three sentences of reasoning are filled in from these numbers by a fixed template. The same inputs always give the same JSON, and so the same hash.
Priors
| Endpoint class | Phase 2 | Phase 2/3 | Phase 3 |
|---|---|---|---|
| cognitive-functional | 18% | 16% | 12% |
| neuropsychiatric | 32% | 30% | 30% |
| biomarker | 50% | 50% | 50% |
| other | 22% | 20% | 20% |
The anchor is Cummings et al. 2014: 413 Alzheimer's trials over 2002 to 2012 and "an overall success rate during the 2002 to 2012 period of 0.4% (99.6% failure)". That is the rate of reaching approval, a higher bar than one endpoint, and it is a 2002-2012 figure.
The bond
The agent bonds 2,000 $PLCB on each open market, from the house address, in the contract. After finalization the bond comes back in full if its Brier is at or under 0.25; above 0.25 half is burned. Until $PLCB is set on the contract, forecasts are hashed and scored, and the bond is posted from the first night after.