14 stored builds since 2026-09-06 — every numeric leaf of this panel is written to signal_history at each build, so a repeated value can mean the upstream had not published, not that nothing moved.

A sparse statistical jump model: k-means clustering with an explicit penalty on switching states, so the label cannot flap day to day the way a plain hidden Markov model's does. Every day is labelled by online inference — using only data available up to that day. We do not publish an in-sample regime chart, because a curve fitted with hindsight describes the past rather than forecasting anything.

13 stored builds since 2026-09-07 — every numeric leaf of this panel is written to signal_history at each build, so a repeated value can mean the upstream had not published, not that nothing moved.

Every forecast carries a hard, machine-checkable threshold and a deadline. At the deadline an official series decides it — there is no judgement call and no quiet deletion. Signals that make a mean-reversion or persistence claim are snapshotted and scored 28 days later on the same terms (90 states recorded, 0 scored so far).

A 28-day score means nothing if the model behind the claim can be refitted, inside those 28 days, on the window it is about to be scored against. So every scoreable claim now takes a sealed holdout over its own forward window: the seal carries a hash of the recorded claim, so the claim cannot be reworded afterwards, and any read of that window before the horizon raises and is written to an audit table whether or not the caller catches it. 0 taken, 0 opened on schedule, 0 refused reads on the record. Across the tracked signals as a family, Hansen's test for superior predictive ability asks whether the best of them beats a coin at all once you account for having looked at every one: not yet computable, because it needs two tracked signals with at least 8 scored claims each; 0 qualify so far.

16 stored builds since 2026-09-04 — every numeric leaf of this panel is written to signal_history at each build, so a repeated value can mean the upstream had not published, not that nothing moved.

Our research loop lets an optimiser propose signal rules, but the code that grades them is immutable and the most recent years of data are held back where the optimiser cannot see them. A rule is only kept if it improves on data it was never shown.

Cumulative abnormal return after each trigger, measured against the sample's own drift, with 95% confidence intervals from 2,000 bootstrap resamples. Only results whose interval excludes zero are listed — and the sample size is shown beside each, because a confidence interval on seven events is a hint while one on fifty is a finding.

36 headline series across 10 desks, each read four ways: level against its own trailing mean, position in its full history, calendar-month seasonality, and an empirical 63-day range. The most stretched is Copper at the 100th percentile of 415 observations back to 1992-01-01. Only 3 of 36 show a calendar pattern larger than their own noise (Gasoline (Gulf), 2s10s slope, 5y breakeven) — the rest are shown with that stated, because a seasonal chart drawn on noise is the most convincing bad chart in finance. Behind the tabs: 18 distinct series decomposed with statsmodels 0.14.6, 18 with a two-state Markov filter and 18 with an analog episode table scored against a linked price.

What changed since last week

cumulative abnormal return vs the sample's own mean drift, aligned by DATE to the nearest session; only the first day of each episode counts as an event; 95% CIs from 2,000 bootstrap resamples. Events are rebuilt from the underlying official series, so these are historical relationships rather than a record of calls we published. Window 10 sessions before, 20 after. As of 2026-09-19 13:12 UTC. Floor for a conditional claim: 20 episodes (grade_t_thresholds.json).

A fundamental exists on the day a quarter ends and becomes public on the day the filing is submitted, a median of 31 days later in this universe. A backtest that stamps the number at the period end therefore trades on it for about a third of every quarter before anyone could have seen it, and is paid for the price move that happened while the market was still waiting. Four of the five quantamental repositories this work draws on do exactly that and none of them measures it. So we run the same earnings-yield quintile long-short twice over 147 names and 155 weekly rebalances, identical in every respect but one: which vintage of the fundamental each arm may see.