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>
"Market Regime" and "Regime Monitor" sat next to each other in Admin -> Jobs
(pipeline steps 4 and 5) reading as the same job. They are unrelated, and the
names had it backwards: "Market Regime" is the SPY 50/200 guard that drives the
TopBar trend dot and the counter-trend warning on setups, so it changes what a
setup shows; "Regime Monitor" is the observational AI/Tech thermometer that
explicitly feeds no trades. The more consequential job had the vaguer name.
market_regime "Market Regime" -> "Market Trend (SPY)"
regime_monitor "Regime Monitor" -> "AI/Tech Risk Monitor"
Display strings only. The job *ids* are persisted -- they key the pipeline step
list, cron config, runtime tracking and run history -- so they are untouched,
as is the /regime route, which keeps existing links working.
The label the admin UI renders comes from JOB_LABELS in admin_service (via
routers/jobs.py), not from the scheduler's APScheduler `name=`. Both are updated;
only the former is user-visible.
Carries the vocabulary through the rest of the surface so it does not half-land:
page title, nav ("Regime" -> "Risk"), the empty-state instruction that names the
job to run, the quadrant alert toggle, the morning-pipeline hint, and the
Telegram alert headline ("Regime quadrant change" -> "AI/Tech risk quadrant
change"). No test asserts any of these strings.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two review findings on 46ace50.
[P1] Raising HY_OAS_WINDOW_DAYS to 700 only reached newly computed rows. A
routine run recomputes the latest trading date alone, and `rebuilding` was keyed
on "no v3 snapshot exists at all", which is false once the cutover has run --
so every row already written kept the credit gap the wider window exists to
close, indefinitely.
Adds SENSOR_REVISION: stamped into each snapshot, absent on pre-marker rows
(read as 1), and a stored revision below the current one triggers exactly one
reseed. Deliberately not METHODOLOGY, which would partition the history API and
discard the cached event study -- neither warranted, since the study recomputes
its Warning series from source rather than reading snapshots and so cannot be
staled by a reseed.
The reseed is bounded by REBUILD_LOOKBACK_DAYS in calendar days rather than a
session count, because the binding constraint is the OAS fetch: each replayed
row needs W3's 20-business-day lookback inside HY_OAS_WINDOW_DAYS. Replaying by
session count would have left the oldest stored rows unrepaired -- the exact
rows the fix targets. At 672 days the replay covers ~464 sessions, W3's oldest
requirement lands on the first fetched OAS day, and the ~400-session series the
cutover wrote is fully covered. A test asserts that relationship so the two
constants cannot drift back into recreating the gap.
[P2] With nothing ever collected, current_observation returned available=true
and the default placeholders -- "unknown" for every hyperscaler, "mixed" for the
reaction -- so the card announced a reading that never happened. Those are the
absence of an observation, not an observation of absence. Gated on `observed`
(non-null fetched_at, the one field every path writing real content stamps),
which blanks the content and drives a proper empty state naming where an admin
collects one. This was a regression from 46ace50; fundamental_overlay never had
it, since no observation means no effective date means pending.
Also renames the leftover v2 identifiers in the touched paths
(rewrite_existing_v2, latest_v2).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The page had twelve stacked blocks, several of them different views of the
same numbers. The quadrant plot and the score-history chart drew the same two
series from the same query key, which read as two datasets; they are now one
card with a Time | Path toggle. The two pillar disclosures become one grouped
table, and three prose blocks (data quality, basket, coverage) become one
provenance chip strip. Page text is now limited to what changes how the reader
interprets today's number; the rest moved to the methodology doc.
Removes three stale-threshold bugs of one class. The quadrant fell back to v2's
60/60 dividers when quadrant_config was absent -- the real values are 50/40 and
they feed alert_service, so the chart could disagree with what actually fires.
The gauge fell back to v2's 30/60/80 band ticks, and drew a divider line that
always landed on its own "elevated" tick. The time series' reference lines were
at 30/60/80, which correspond to nothing in v3; they are now per-axis dashed
lines read from the same quadrant_config. Rendering also surfaced a live
clipping bug inherited from the old chart: margin.left -18 against YAxis
width 28 left ~10px for a 3-digit label, so every Y tick was cut off.
HY_OAS_WINDOW_DAYS was 400 *calendar* days while a rebuild replays
REBUILD_SESSIONS = 400 *trading* sessions (~579 calendar days), so the oldest
~180 days of any rebuild got no OAS at all and both credit sensors returned
None. State then lands at 80% coverage and Warning at exactly MIN_COVERAGE, so
both still publish bands -- a series that looks homogeneous while its oldest
rows were scored without credit. Widened to 700. This needs no methodology
bump: C1 reads [-1] and W3 reads [-21], both from the end, so widening only
prepends and every live score is bit-identical. Sequenced deliberately, since
acting on the open findings below bumps METHODOLOGY and fires the rebuild.
A just-collected fundamental observation was hidden until its effective date --
one day, three over a weekend -- because the live reading called the
point-in-time function, so refreshing appeared to do nothing. That was the
opposite of what the doc claimed. fundamental_overlay stays the gated record
(it runs for every replayed date during a rebuild); current_observation is the
live reading and reports the effective date instead of blanking the content.
Nothing in the overlay is scored, so showing it early cannot reach a published
number.
Documents four calculation findings. Three are not implemented, since each
changes a published score and so requires a v4 cut: State's top band is a
credit-event band (credit returns 0.0 rather than None below the 3.5 anchor, so
it is pinned at zero at weight 20 -- with everything else pegged State computes
to exactly 80.0, the breaking threshold); V1 saturates at VIX 30; and the
deliberate max(P1,P2,P3) defeats P3's anchoring because P1 is binary.
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 combined score collapsed two distinct signals into one not-very-meaningful
number. Replace its gauge with a quadrant scatter that shows both axes directly:
x = regime index (coincident), y = early warning (breadth divergence), with a
trail of the last 60 sessions and today highlighted.
The four quadrants make the readings legible — ① hot & brittle (narrow melt-up,
shakeout risk), ② transition, ③ healthy & broad, ④ real downturn — and the trail
surfaces the actual tell: the ①→④ move (early warning rolling over as the regime
index climbs = divergence resolving downward). Combined still shows as a line in
the score-history chart. Frontend-only; reuses the history endpoint. Lazy-loaded.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Plots the index, early-warning, and combined scores over time beneath the live
gauges, with a 1M/3M/6M/All range toggle and band reference lines — so the trend
and any divergence between the scores is visible, not just today's snapshot.
- Backend: GET /regime/history + get_regime_history (the three scores per
snapshot date from regime_snapshots).
- Frontend: recharts line chart, lazy-loaded so recharts ships in its own
regime-tab chunk instead of nearly doubling the main bundle.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The "warned a median -18 days later" line was the median-over-1-event trap: the
coincident baseline's 60d median is a single lucky event, while breadth warned on
7. Replace it with the honest coverage framing (7/11 vs 1/11) and flag that the
median-lead comparison is unreliable when coverage differs this much.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The event study showed the breadth-divergence signal genuinely leads (warned
before 7/11 drawdowns, ~6 weeks median, where the coincident baseline almost
never did). Surface it live to observe before deciding how to embed it — kept
separate from the index, not folded into its weights.
- regime_monitor daily job now computes breadth-divergence live and attaches a
separate early_warning score plus a combined blend (weighted mean, default
0.6/0.4, configurable via combined_weights) to each snapshot, including the
backfill so the 7/30-day trends populate immediately. Stored in breakdown_json
— no schema change. Best-effort: a breadth failure can't break the index.
- get_regime_monitor returns the index, early_warning, and combined scores each
with 7/30-day deltas.
- Regime tab shows three gauges (generalized ScoreGauge): coincident index,
early warning, and a compact combined blend. Stale snapshots render "—".
Note: the daily regime job now also does a universe-wide breadth scan.
Co-Authored-By: Claude Opus 4.8 <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>
A new /regime tab scoring how far the AI/Tech bull regime has deteriorated
toward a re-rating as a single 0-100 index with per-signal breakdown and a
7/30-day trend. Intentionally decoupled: nothing reads its output to gate or
score trades — the daily-pipeline membership is scheduling only.
- regime_monitor_service: price sub-scores (P1-P6 via Alpaca, like
market_regime), VIX + HY credit spreads via a small FRED helper, weighted
aggregation over available signals (missing source -> n/a, dropped from the
denominator), one snapshot row/day, and a ~90-day history backfill by
replaying the already-fetched series as-of each past day.
- F1/F3 fundamentals proposed by the configured grounded LLM (reuses
sentiment_provider_service config resolution), with a manual override + lock.
- regime_snapshots table (migration 011); endpoints on the existing market
router; admin-editable weights/threshold; standalone /regime page.
Data needs: prices via Alpaca, VIX/credit via FRED (optional key — signals show
n/a without it). No LLM needed for history.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
