The TPG Cutoff Method

01

Fix the evidence cutoff

Every forecast states the date its evidence stops. We use only sources published on or before that date, and each source is listed with its publication date so anyone can check. No hindsight, no leaked documents, no private or client information, no material non-public information.

The cutoff is what makes a forecast honest. Anyone can explain an outcome after it happens; the only test that counts is what a careful analyst would have concluded with what was knowable at the time.

02

Psychology: what do the decision makers need?

Politics is decided by people under pressure. We start with their incentives: the audiences they answer to, the cost of backing down in public, what they cannot be seen to lose, and their record of following through on threats and promises versus bluffing.

We describe behavior on the public record. We never diagnose anyone or speculate about private mental states.

03

History: find the reference class and the closest parallel

Before judging the specific case we ask how often this kind of thing has happened: how many funding deadlines ended in a shutdown, how many ceasefires held, how often an incumbent's party lost seats. That reference class gives the base rate.

Then we name the closest historical parallel and, just as important, exactly where it breaks. A parallel without its breaking points is a story, not evidence.

04

Statistics: base rate, adjustments, one number

We start from the base rate and adjust in the open for what is different this time, showing the arithmetic in words. The result is one probability with an honest range, broken into scenarios that add up to 100 percent.

We compare our number with prediction market prices when a market exists and explain the gap. Agreeing with the market is not the goal; being right and calibrated is.

05

Read between the lines

Statements are cheap; troop movements, spending, evacuations, resignations and cancelled ad reservations are expensive. We weigh costly signals over cheap talk, note what an official readout conspicuously left out, and track personnel moves, sequencing and shifts in tone.

06

Trace every statistic

Every headline number on a forecast page cites a source, and the source entry carries the verbatim sentence the number came from. An automated check fetches every source and confirms the quoted sentence is on the page before a forecast is published.

07

Lock it, update by version, score it in public

Each published version is locked with a sha256 hash of its full contents, and the build refuses to deploy if a locked version changes. When the evidence moves we publish a new version beside the old one, with a new cutoff.

When a question resolves we score it with the Brier score (0 is perfect; a permanent 50 percent guess earns 0.25) on our final call and on the time-weighted average of every version, so a late correction cannot hide a long stretch of being wrong. Hits and misses stay on the track record.