ruff F401 failed the deploy: `import pytest` in test_event_study.py outlived the
pytest.approx assertion it was added for.
Compiling the migration for Postgres while checking that turned up a second
defect worth fixing while the table is still empty. It created a unique
constraint *and* a plain index on effective_date, while the model declares
`unique=True, index=True` -- one unique index. Both enforce uniqueness, but the
pairing left a redundant second index on the column and a permanent diff for
autogenerate to keep trying to reconcile. Now renders byte-for-byte what the
model declares, matching RegimeSnapshot.date.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The Warning study measured a fitted percentile crossing that nothing consumes.
What reaches Telegram is a quadrant change: fixed 50/40 dividers, hysteresis,
two-session confirmation, 3-day cooldown. Those thresholds are constants, not
fits, so there is no training set to protect and all 11 detected corrections are
evaluable instead of the 4 that fell in a holdout.
Replaying it: 1/10 corrections, 0.9 false alarms/year. Random alarms at the same
firing rate match or beat that in 65% of draws. The panel now carries ablations
(does the quadrant machinery earn its place?), external baselines (does the score
earn its complexity?), and that null, because a bare "2 of 4" was unreadable in
either direction. Nothing in the alert path was retuned on the strength of it.
Fundamentals become a third channel rather than a term in either score. v3 cut
them arguing 12+8 of 100 points "could not change any published conclusion" --
true only when every technical sensor reads zero; weighted they moved the bar for
the 40 divider from 40 to 25. But no fusion weight is measurable either: with ~10
events and no fundamental history, any weight is a policy preference presented as
a measurement. So the read is a categorical state (supportive/neutral/adverse/
unknown) with an evidence grade, derived by fixed rules from stored facts, read
by confluence. The LLM extracts and explains; it does not score.
Absence stays absence throughout. `unknown` is unreachable by averaging, a stale
or empty observation may display but never confirm, extraction failures map to
`unknown` rather than `mixed`, and the study rows are coverage-matched and marked
not-measurable until enough corrections are covered -- otherwise a fortnight of
observations renders as 0/10 and reads as a failed test.
Observations become a real time series (migration 033); they lived in a single
overwritten settings slot, so no history existed to replay. Pre-rename snapshots
are adapted rather than discarded. METHODOLOGY stays v4 -- no score changed --
so no reseed; STUDY_SCHEMA moves to 3 and discards the cached report.
Post-deploy: re-run Event Study from Admin -> Jobs. The panel reads "not run yet"
until then.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The v3 cutover run scored 2/4 corrections warned against v2's 3/4, which reads
like a regression and is not one. Only 4 of the 11 detected corrections fall in
the holdout, so recall is one event from a different headline -- and the event
that flips is decided by threshold placement, not by what the score saw. "v3
without the credit sensor" catches 2025-02-21 at a *higher* threshold (35.5)
than shipped v3 misses it at (32.3), because the alarm rule needs a rising edge
and a lower threshold can fire outside the horizon then never reset below.
Two caveats are now computed and surfaced rather than left for the reader to
infer:
- Holdout event count against MIN_EVENTS_FOR_CONFIDENCE. The summary sentence
states how many of the detected corrections actually fall in the test period.
- Warning-sensor coverage across the split. The score renormalises over what is
available, so a training window predating a sensor's history freezes the
threshold on a different construct than the holdout is measured against. At
the cutover that is 39% of training sessions with all three sensors versus
100% of the test period, credit history beginning 2023-07-25.
Restricting the threshold to sensor-matched training sessions was tested and
rejected: those sessions are a calm recent stretch, so the threshold falls from
32.3 to 22.5 and false alarms rise from 3.3 to 8.6/yr. It swaps a coverage bias
for a regime-selection bias. The report states its limits instead.
_warning_series now returns per-session sensor counts alongside the scores.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The LLM-sourced capex/earnings observations carried 12+8 of 100 Warning points,
so both pegged at 100 produced a Warning of 20.0 -- below the event study's 25.3
alarm threshold and still inside the "stable" band. The reading was
arithmetically incapable of changing anything on screen, which is why refreshing
it appeared to do nothing. They are now a qualitative overlay reported beside
the scores rather than diluted into them.
Calibrated against the 408 v2 sessions to 2026-07-24, reproduced offline from
Alpaca + FRED; the harness matched the stored prod distribution exactly before
any parameter was changed.
State:
- P3 used dd_pct * 5, reaching 100 at a 20% drawdown -- the 90th percentile of
the observed distribution -- so 39/408 sessions sat at exactly 100 with no
resolution left during the part of a selloff that matters most. Replaced with
anchored breakpoints keeping headroom past the observed 36% maximum, blended
2:1 like P1/P2 instead of max(). P3's realized share of State falls from 65%
to 40%, matching its nominal weight.
- Credit level is now anchors-only. ICE capped FRED's BAMLH0A0HYM2 at a rolling
3-year window in April 2026, silently turning the 10-year percentile leg into
a 3-year one that scored 20 points of stress at an OAS of 3.5 -- the level its
own anchors call "mild". The anchors already encode the long-run distribution.
Warning:
- Added HY OAS 20-session widening (25%). The level is pinned at zero below the
3.5 anchor; its rate of change is not.
- Divergence tapers to a 0.35 floor instead of a hard price_ret >= 0 gate, which
zeroed the sensor through every decline: on 2026-07-24 the basket shed 10
points of participation in 20 sessions and Warning printed exactly 0.
- The event study and the live monitor now share one sensor definition, so they
cannot silently drift apart.
Bands are per axis (State 20/50/80, Warning 20/40/60) with quadrant dividers at
50/40; v2 Warning never exceeded 64.9 against a shared 60, leaving that half of
the quadrant unreachable. Realized shares: State 73/15/8/3%, Warning 69/20/8/3%.
Snapshots now record credit_history_days and vix_history_days -- the percentile
defect went unnoticed for months because nothing asserted the window the code
claimed.
Cutover: the first run rebuilds 400 sessions automatically; the Event Study job
must be re-run, as its cached report self-invalidates on the methodology check.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The first run gave only 2 events (N=2 is anecdote, not evidence) and an unfairly
weak coincident baseline, so the +42d lead couldn't be trusted. This makes the
measurement meaningful:
- More, cleaner events: default drawdown threshold 15%→10%, and dedup switched
from "recover to the high" to a rising-edge + cooldown (40d), so distinct
drawdowns each register instead of merging.
- Fair comparison: each indicator now warns at its OWN 80th percentile instead of
a shared absolute 60, removing the artifact that muted the coincident baseline.
- Per-event breakdown (date · depth · breadth lead · coincident lead) so a median
over a tiny sample can't hide an apples-to-oranges comparison — you see whether
both warned on the same drawdown.
- Surface precision/recall (best row) + base rate per indicator — the honest edge
read, not just lead time.
Re-run the Event Study job to regenerate the cached report in the new shape.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds a leading-by-construction candidate and the harness to measure whether it
actually leads regime breaks, before any of it earns weight in the live index.
- breadth_service: % of the stored universe above its own 200-DMA + a divergence
score (benchmark price up while breadth falls, nudged by low breadth). Genuinely
leading because it keys on divergence, not level. Not wired into the live score.
- event_study_service: detect drawdown events on the benchmark, then measure each
indicator's median lead time (event-centered) and precision/recall vs. the base
rate (signal-centered). Compares breadth-divergence against the deterministic
coincident price composite (reuses the regime price sub-scores). Price/breadth
only — reproducible, no LLM/FRED.
- Manual "Event Study" job (Admin → Jobs), GET /regime/event-study, and an
inline early-warning panel on the Regime tab with an honest small-sample caveat.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
dennisthiessenandClaude Opus 5