The FOMC rules, read forward
two rules fixed on 2026-09-03, one row per decision day since, added after the session
Key finding. Two directional patterns turned up after the FOMC study's list of comparisons was closed: the drift into the press conference reverses when it was a dip, and the last hour of a decision day sells off. Fifteen looks produce one p of 0.02 by chance, so neither was claimed. Both are fixed here as rules and scored on every meeting from 2026-09-16 on, one row per decision day, added after the session and never edited.
The FOMC study ran 36 comparisons it had listed in advance and then, at the owner's request, fifteen directional splits it had not. Two of the fifteen stood out. Fifteen looks is roughly the number at which one p of 0.02 appears by chance, so the study printed both as hypotheses and promised a forward test. This page is that test. The rules below were fixed on 2026-09-03; the feature store appends each decision day on its own the night it trades; a row is added here after each meeting, computed by a script from the store, and never edited afterwards.
The two rules
| when | condition | trade | scored as | |
|---|---|---|---|---|
| rule 1 | 14:30 on a decision day | spot fell over the previous 15 minutes (ret15 ≤ −0.5) | long | spot at 15:00 minus spot at 14:30, points |
| rule 2 | 15:00 on a decision day | none | short | spot at 15:00 minus spot at 15:30, points |
In sample, for scale and nothing more: rule 1 fired on 17 of 35 meetings and averaged +21.6 points, up on 14 of 17, t = 3.2; rule 2 averaged +8.3 on 35, up on 19. Those numbers were found by looking; the column that counts is the one below.
The ledger
| meeting | drift 14:15 → 14:30 | rule 1 | 14:30 → 15:00 | rule 1 running | rule 2, 15:00 → 15:30 | rule 2 running |
|---|---|---|---|---|---|---|
| 2026-09-16 | the first reading | |||||
Drift is the 15-minute move into the press
conference. Rule 1 trades only after a dip; a rising drift is a row with no trade. Running
totals are points per rule, with the count of winning rows. Each row is written by
pipeline/event_stats.py on the production box, in the same six-hourly job that
rebuilds the market calendar, from the day's own one-minute spot series in the public
archive. No one adds, edits or approves a row.
What would count
Rule 1 has about four readings a year. If after eight of them the running total is positive and at least five of eight are up, it has matched its in-sample shape and earns a second study on the full record; if it is negative, the post that found it gets a line saying so, and the rule is retired here in public. Rule 2 reads every meeting; sixteen rows with a positive total and a t above two on the day-clustered error would move it from hypothesis to finding. Anything short of that stays a hypothesis, whatever the last row says. Nothing on this page is investment advice; see the Terms.
Part of FirmTape research. Measured on the free session archive; every session is free to replay.