How We Think

Most biotech research is non-numerical. KOL panels and analyst notes describe a drug's prospects in adjectives. Where numbers do appear — crude base rates, broad class averages — they tend to give everything in a category the same odds. The result is research that cannot be wrong in any measurable way, and therefore cannot get better.

Regatta starts from the opposite commitment: a forecast is only worth making if it can be scored. Three ideas follow from that.

1 — Calibration

Confidence has to match frequency.

A forecaster who says "70%" should be right about seventy percent of the time when they say it. Calibration measures exactly that, across many resolved events. It rewards honesty about uncertainty and penalizes overconfidence — the failure mode that makes a confident wrong call so expensive.

2 — Decomposition

Break the event into parts you can actually estimate.

"Will this drug be approved?" is hard to answer directly. The component questions — does the endpoint hold, is the safety profile clean, how has the agency treated this mechanism before — are tractable. Decomposing a catalyst into estimable parts and recombining them produces a probability you can defend and inspect, rather than a number pulled from intuition.

3 — Selection

Let resolved results decide who forecasts.

Track records are earned on events that have actually resolved. Analysts who forecast well are rewarded and continue; those who do not are removed. The population of forecasters improves over time because the mechanism, not a manager, decides who stays.

The work is steady, slow, indexable, and checkable. We would rather publish one defensible retrospective than ten confident predictions.

What this looks like before launch

Until launch, our published work is retrospective and methodological: what predicted past readouts in hindsight, how advisory committee voting has moved over a decade, and decomposition worked through on resolved historical cases. You can find it in Research and in the Regatta Brief.