The pm1 verdict is promising-but-underpowered (n=21): the forward moves are real (+11-14%) but the
sample is too small to conclude. The live loop ingests premarket bars every trading day, so this
automates the "accumulate more days then re-validate" step:
- research/premarket_eval.py: rebuild pm1 from the grown DB, append a timestamped metrics row
(n, days, +1R rate, Wilson-95 LB, median upside/adverse, big-gap +1R, pass/fail) to
data/premarket_eval_history.tsv, and print the current validate output. Watch n climb + the
Wilson bound tighten until the locked bar is decisively cleared or failed.
- deploy/systemd/momentum-premarket-eval.{service,timer}: weekday-evening (Mon-Fri 20:00) oneshot,
Persistent=true (catches reboots). Installed to ~/.config/systemd/user and enabled.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Every RTH selection/entry lever has now been refuted (6 of them separate a population but
don't convert through the entry). The one axis never touched is PREMARKET — the source
trader's actual edge (his words: news drops pre/after-hours, NO circuit-breaker halts pre-market
so moves run clean, and RTH after 9:30 is choppy/unpredictable — the exact window our bot trades).
strategy/evaluation/premarket_setup.py::detect_premarket_setup — pure, operates on PREMARKET bars:
gap vs prev_close + real premarket volume floor (the DXST-vs-CF tell: low pm volume = fail) +
price floor + fresh premarket-high breakout + premarket-VWAP hold. Calibrated to the thin pm tape
(cum pm volume p50 ~12k shares, so a 20k floor catches the genuine movers not 1%-gap noise).
Real-tape validated + SELECTIVE: 6/30 fires JEM(245%)/CELZ(143%); 7/1 fires LHAI(166%)/SOC/SNDQ —
2-3 names/day (~8% of pm names), vs the liquid detector's 50%. 6 unit tests green.
labeler pm1 track (build/report/validate-premarket): slides the detector over premarket bars,
scores forward on PREMARKET bars via the same _forward_labels (the trader enters+exits pre-9:30),
stratifies +1R by gap. Pre-registered bar locked 2026-07-01 (median pm upside >=+10%, adverse >-6%,
+1R >50%, big-gap subset clears). SHADOW-ONLY. Real caveats it sets aside: thin premarket fills, and
live premarket needs extended_hours execution the bot doesn't have. Full 26-day backtest building.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The systemd user service runs with a minimal PATH that excludes /usr/bin, so yt-dlp's audio
postprocessor reported 'ffprobe and ffmpeg not found' and every video silently fell back to
captions. Pass ffmpeg_location=config.FFMPEG_LOCATION (shutil.which with /usr/bin fallback) to
yt-dlp so the Whisper path works under systemd. Verified: yt-dlp download + ffmpeg extract now
produces a wav. Reset the 4 caption-fallback videos to pending so the backlog gets Whisper.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Extraction fix: the anti-hallucination check (a claim's verbatim_quote must appear in the
source window) was too brittle against auto-captions — fragmented segments joined with
newlines meant a valid quote spanning a join was dropped (0 claims on a 623-segment captions
transcript). Now whitespace-normalized matching (_norm + _quote_supported, ~40-char prefix):
the SAME transcript yields 6 real claims (WETG +160% reverse-split+news; 4 AURC trade claims
with prices; a rejection concern), each quote verifiably present. Whisper output is cleaner
still; this just makes the captions fallback usable too.
- tests/unit/test_youtube_extraction.py: 5 pure-helper tests (norm, quote-support across joins,
hallucination rejection, verbatim-quote-mandatory drop, chunk time-spans).
- youtube_claims/momentum-transcripts.service: reboot-surviving systemd user unit (Restart=always,
loads secrets from .env; Whisper option A/CPU, env-flippable to B/GPU).
END-TO-END VERIFIED (captions path, no ffmpeg): detection -> claim -> transcribe -> extract ->
6 claim rows in Postgres. Full Whisper path needs `sudo apt install ffmpeg`, then
`systemctl --user enable --now momentum-transcripts` to process the 11-video backlog at
Whisper quality + poll for new uploads.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A descriptive ingestion layer feeding the downstream reasoning model: monitors YouTube
playlists, transcribes new uploads, and extracts structured, verifiable CLAIM records
(with provenance + a mandatory verbatim quote). It records WHAT was claimed; it never
judges correctness or tradeability -- all judgment stays downstream, in one place.
Built per the operator's spec. Two clocks coordinated through Postgres (§2): a 30-min
detection poll + a paced concurrency-1 transcription/extraction worker. Reuses the app's
infra: shared Postgres (isolated `transcripts` schema), the app's Ollama for extraction
(free, on-LAN, private -- transcripts never leave the network), the config.py host resolver.
youtube_claims/schema.sql playlists / videos / claims (§7-8); claims.verbatim_quote NOT NULL
youtube_claims/db.py psycopg2; claim-by-atomic-status-flip under FOR UPDATE SKIP LOCKED
youtube_claims/detection.py playlistItems.list over urllib; dedupe by video_id (§5)
youtube_claims/extraction.py local Ollama -> descriptive claims; drops any claim lacking an
exact verbatim quote present in the source window (anti-hallucination)
youtube_claims/transcription.py yt-dlp -> faster-whisper (medium/cpu/int8, option A;
initial_prompt=watchlist) with youtube-transcript-api captions fallback
youtube_claims/worker.py paced drain; per-video state machine + retries/backoff
youtube_claims/run.py APScheduler poll + worker loop in one process; CLI for setup
VERIFIED LIVE: schema applied; playlist "Anatomy of a Trade" (Ross Cameron) added; poll
ingested all 12 videos (1 API unit); Ollama extraction pulled clean descriptive claims
(TC/DXST/CF) with exact verbatim quotes from a real transcript. Whisper transcription pending
ffmpeg (`apt install ffmpeg`, needs sudo) + the yt-dlp/faster-whisper deps (installing via uv).
Secrets (YOUTUBE_API_KEY) live in gitignored .env only. Whisper is option A (CPU); env-flippable
to option B (LAN GPU large-v3) later.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Mirrors the runner track: slides detect_liquid_intraday_breakout over RTH history, emits
non-tradeable liq1 setups scored by _forward_labels, stratifies +1R by rvol + cumulative
dollar-volume. Pre-registered bar locked 2026-07-01. Early in-sample read (5 days, 1511 fires):
+1R 51% beats ORB 35% on same days, but median 60m upside only +0.6% (tight-stop artifact);
rvol/liquidity strata flat. SHADOW-ONLY. Full 26-day out-of-sample build pending.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The 2026-07-01 forensic + the source trader's day-14 notes point at a thesis-scope gap:
the day's only clean, held, TRADEABLE +2R moves (MSTR, SSPC) were LIQUID names that ignited
intraday with ~zero premarket gap -- invisible to the premarket-gapper universe -- while the
cheap "held trenders" were untradeable dust and the liquid gappers the bot chased whipsawed.
strategy/evaluation/liquid_breakout.py::detect_liquid_intraday_breakout -- a pure detector that
makes LIQUIDITY (cumulative + per-bar dollar-volume) a first-class gate (the thing runner.py
lacked, which let it fire on dust), plus a price floor + fresh-session-high breakout + VWAP hold.
NO gap, NO blue-sky, NO float. Validated on the real tape: fires on MSTR (@bar15) and SSPC (@bar90),
correctly silent on the sub-$2 dust (DXF) and the <$1M single-print names.
Design finding baked in: RVOL is RECORDED, not a hard gate. MSTR/SSPC were NOT within-session
volume surges -- MSTR was liquid from the open (recent/early vol ratio ~0.5-1.0 the whole grind),
so a hard rvol>=3x "surge" gate structurally kills the steadily-liquid grinders this lane exists to
catch. rvol is stored as a stratification feature (the shadow track will measure whether high-rvol
fires convert better, like the runner GRADE); it only gates if rvol_min>0 is passed.
Selectivity is broad (~50% of the liquid universe fires at least once/day) -- this is a candidate
POPULATION definer, not a precise signal. The decisive question (does it convert to +1R better than
the ORB, and does any rvol/liquidity stratum beat it) is for the shadow track + a pre-registered
forward bar. SHADOW-ONLY: never routes to approve_order; 6th selection signal, must prove conversion
after cost before any live gating. See memory thesis-scope-liquid-intraday-momentum.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root-caused (workflow wjy4nzpbw, adversarially verified) why the live loop's per-cycle
log printed `execute: ... auto_executed=1` ~951 times after 09:45 ET on 2026-07-01 with
ZERO matching broker orders: reject_order() returns a truthy {"ok": True} dict that
process_auto_approvals() appends unconditionally, and step_execute logged len() of that
list as auto_executed. So every daily-entry-cap rejection masqueraded as an execution,
hiding the fact that the bot was hard-capped at 5 fresh entries for the rest of the day.
This is the log-VISIBILITY half only (safe, zero trading-behavior change):
- reject_order() tags its return rejected=True + reason, so a rejection is distinguishable
from a real execution (both carry ok=True).
- step_execute() counts auto_executed = results with a real broker_order_id, and appends
`auto_rejected=N (reason x N, ...)` so a tripped cap is VISIBLE instead of silent, e.g.
`auto_executed=0 auto_rejected=1 (daily_entry_capx1)`. ASCII-safe for the systemd log sink.
NOT included (deferred): the behavior fix (cap counts accepted submissions incl. unfilled
resting limits that get backed out, never decrements — so 4 phantom orders burned the whole
day's entry budget). That change lets the bot take more real shots, which given the live-
confirmed entry-conversion weakness likely means more churn — pair it with entry-quality
work, don't ship solo. See memory daily-entry-cap-counts-submissions.
Takes effect on the next loop boot; deliberately NOT restarting mid-session (the cap counter
is in-memory, so a restart would reset it to 0 and re-open entries into today's tape).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fidelity pass against the leading-gainer-runner workflow (wgcbvb3ce) synthesis:
- runner_quality.py: reorder calculate_runner_quality signature to the spec's
(gap_pct, vol_burst, velocity, above_vwap, ...) — behavior identical (all callers
use keywords), now matches the documented signature so a positional caller reading
the spec is correct.
- labeler.py: close build_plan step (3) "attach the runner grade". The grade is a pure
function of components already persisted on each fire (gap/velocity/burst/vwap), so
it is recomputed exactly at report/validate time (catalyst-OFF) — no 34-min rebuild.
report-runner now stratifies forward +1R and median 60m upside BY GRADE; validate-runner
adds the A/B-subset test (does concentrating on top grades salvage the locked bar?).
FINDING (385 fires, 25 session-days): the detector still FAILS all 5 pre-registered
bars (the workflow predicted this — SVRE/JEM put 100% of the move pre-market, RTH fires
confirm at/after the top). BUT the grade is MONOTONE with forward outcome:
grade A: +1R 45%, median max-upside 60m 17.7% (n=73)
grade B: +1R 41%, median 9.2% (n=186)
grade C: +1R 35%, median 7.8% (n=106)
grade F: +1R 30%, median 3.5% (n=20)
Grade A ALONE clears the +15% upside bar (17.7%) but not +1R>50% (45%): A-runners have
real continuation, the naive last-close entry still whipsaws. Confirms the grade earns
its keep as a RANKER/sizer (concentrate on A) while the bottleneck stays ENTRY-conversion.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The default ORB grade (quality.py) pins catalyst RUNNERS at F/C: its two heaviest
components structurally break on a vertical spike -- STRUCTURE (30%) reads 0 (no
pullback to classify) and adv-RVOL (25%) reads 1.0x (NULL rolling-20d on fresh
reverse-split names). On the 2026-06-30 tape the source trader's winners (SVRE
+247%, JEM +348%, CELZ +472%) all graded C 0.565, indistinguishable from the bot's
F-grade chop. Two parallel, shadow-only artifacts fix this without touching the live
ORB path:
- strategy/evaluation/runner_quality.py::calculate_runner_quality -- a PARALLEL
grade that drops the two broken components and scores the traits that actually
separate runners from chop (gap saturating at +100% not +10%, price velocity,
own-name volume burst, above-VWAP, catalyst). Validated on the real tape:
CELZ/JEM/SVRE -> B/A, chop (SLS/RIVN/CCXI/BB/NVDX) -> F. quality.py untouched.
- strategy/evaluation/runner.py::detect_leading_gainer_runner -- catches the mover
detect_momentum_ignition is blind to (it requires blue-sky > prior ATH, so it
fires ZERO on beaten-down catalyst pops). Blue-sky deleted; gates = leading-gainer
run + verticality + own-name burst + holding session-cumulative VWAP; PM-inclusive;
tags (never gates) PM-exhaustion. Real-tape in-sample: 5/5 runners fire, 7/7 chop
silent, zero false-fires. Cumulative VWAP is a prefix function, so a precomputed
session_vwap makes a bar-by-bar slide O(n) not O(n^2) (identity verified exact).
- research/labeler.py -- build-runner/report-runner/validate-runner shadow track,
mirroring vwap_reclaim: emit a non-tradeable setup on each fresh fire, score with
the SAME _forward_labels. Pre-registered promotion bar LOCKED before forward data
(median fwd-max >= +15%, median adverse > -8%, +1R > 50%, beat vwap_reclaim,
PM-exhaustion tag must earn its keep). SHADOW-ONLY: never routes to approve_order,
never gates live until the bar clears out-of-sample. The honest open question the
track exists to answer: the detector confirms AT/AFTER the top for PM-driven names
(SVRE fires at +242% already run) -- does that late entry actually convert to +1R?
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Ollama was unreachable — the enrichment/qabot pointed at the 127.0.0.1:30069 netsh tunnel,
which the WSL network no longer routes (same reboot-flip landmine as the DB), while the
direct 192.168.1.5:30068 answers. config.OllamaConfig.from_env now resolves the host: probe
the configured endpoint, else the first reachable known candidate (tunnel 30069 <-> direct
30068 — note the DIFFERENT ports). So enrichment self-heals on boot instead of silently
failing. .env OLLAMA_HOST also repointed at the direct host.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Forensic on 2026-06-30: of the bot's 8 traded names, SEVEN graded F (<0.50) and one B —
it sprayed low-quality chop (SLS F0.12 -$412, RIVN F0.19, BB F0.21, CCXI F0.26 +$374,
NVDX F0.33, NOK F0.48, RDW F0.49; GVH B0.71). The trader made +$616 on SVRE — which the
bot SAW but its 35% gap cap excluded; and even graded, SVRE is F (0.35): low computed
RVOL + no clean pullback structure. So the grade cuts chop but does NOT recognise a
vertical catalyst RUNNER.
Adds ExecutionSettings.min_quality_score (env TRADING_MIN_QUALITY_SCORE, default 0.0 =
off) and a gate in request_approvals_for_ready_signals mirroring the VWAP gate: a signal
scoring below the threshold is skipped and logged rule_type=quality_below; fail-open on a
missing score. 3 tests (skip-low-when-enforced / allow-when-off / allow-at-threshold);
44/44 trading-execution tests green.
Honest scope: this is the "trade less chop" lever (evidence: the bot's own labeled data
has q>=0.55 win ~71% vs q<0.55 ~36%). It will NOT catch SVRE-class leading gainers — the
grade rewards pullback structure the runners lack, and the gap cap excludes them upstream.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Per the operator: a carried position sitting at breakeven shouldn't ladder off its stale
overnight entry. When a held position is at breakeven+ with no real R, the +0.25R trail
now RE-ANCHORS to today's opening range — reference = the session open, R = the first
5 RTH bars' high-low — so the stop tracks today's move from the open and locks in tighter
than the wide entry-based synthetic R. Until the opening range forms (first ~5 min) it
trails off the synthetic fallback R, then re-anchors once (one-shot, tracked in
_open_anchored; cleaned up when the position closes). The stop only ever ratchets UP and
never drops below the live breakeven. 1 test added (re-anchors + ladders tighter than the
synthetic); 11/11 exit-manager tests green.
For KEEL: at tomorrow's open its stop ladders up from the open's range instead of dangling
at the $5.56 breakeven.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Focused review pass on the session's safety-critical additions, applying the safe wins:
- exit_manager: drop the dead `_Tracked.last_stop` field (written, never read) and its
assignment; guard `_rest_protective_stop` against a degenerate qty<1 (a 0-qty submit
is rejected and would churn) + document the cur-falsy->breakeven fallback; reword the
synthetic-R comment (the branch is also hit on a healthy trade once breakeven ratchets
the stop to entry, not only the stripped-stop case).
- exits: surface default_trail_r_pct in ExitConfig.describe() (boot-log observability)
and clamp it to [0, 0.95] in from_env (a >=1 / negative synthetic R would put the stop
at/above entry).
- db_pg: comment the gate-vs-try-order distinction of _LAN_DB_HOSTS (superset incl.
localhost) vs _LAN_DB_FALLBACKS so a future edit can't break the non-LAN-DSN guard.
- momentum_cli: drop an unused `import sys`.
Behavior-preserving; 29/29 exit/exits tests green. Deliberately did NOT extract a shared
flatten-vs-rest helper for the catastrophe/naked blocks — the explicit blocks read
clearer for this safety code and the tests guard against drift.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
KEEL was left at a frozen breakeven stop because its original R was lost — the bracket
stop leg was invalid (the limit filled below the ORB-low, so the stop sat ABOVE entry
and Alpaca cancelled it, which is what left it naked). The manager skipped any position
already at breakeven+, so it never trailed: a winner could only scratch or run
unprotected-above-breakeven.
Now a position first seen at breakeven+ with no real risk distance is tracked on a
SYNTHETIC R (ExitConfig.default_trail_r_pct, default 10% of entry, env
TRADING_EXIT_DEFAULT_TRAIL_R_PCT), so the SAME tested +0.25R step-trail / profit-lock
ratchet the broker stop UP as it rises (only ever up, never below the live stop).
default_trail_r_pct=0 keeps the old frozen behaviour; positions with a real sub-entry
stop are unchanged.
TRAILING is gated on the session being OPEN (one broker-clock check per pass): after
hours there are no new highs to trail under AND a resting GTC stop is often not
replaceable ("accepted" status -> Alpaca 422), so attempting it just churns. The stop
stays protective and ratchets up at the open. 4 tests added (synthetic-R trails up;
disabled=frozen; closed=no-attempt); 10/10 exit tests green.
Live now (after-hours): KEEL is protected at breakeven $5.56; the loop holds that stop
without churning and will trail it up once the session opens.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
KEEL was carried into after-hours as a naked long (its bracket stop leg never
attached — 0 stop orders in its whole history). The enforcement tried to FLATTEN it
with a market order every pass, but a market sell can't fill when the market is
closed, so it churned submit->cancel ~every 30s (76 market-sells / 79 cancels, the
"naked-stop enforced exit KEEL xN" spam) and left the position genuinely unprotected.
Fix (runtime/exit_manager.py): both the naked-stop enforcement AND the catastrophe
backstop now branch on the broker clock. Market OPEN -> flatten (a market exit fills,
unchanged). Market CLOSED -> REST a protective sell-stop instead (GTC) — at breakeven
when in profit, a catastrophe-distance stop below market when underwater — so the
position survives to the open PROTECTED and the next pass sees the stop and stops
enforcing. Catastrophe with an existing (loose) stop while closed just defers to it
(no market churn, no double stop). New client bits: AlpacaPaperClient.get_clock() and
a standalone stop_price on submit_order. 3 exit-manager tests added (open->flatten,
closed->breakeven-rest, closed-underwater->catastrophe-rest); 7/7 green.
Operationally: KEEL was protected at breakeven ($5.56) immediately and the loop now
leaves that stop in place instead of churning.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
From a design pass (CLI spec + dashboard spec + an adversarial glance-safety review)
that read the real output of both surfaces. Principle: each safety/state number appears
ONCE on the first screen, every list has a bounded height capped by relevance (active/open
first), and load-bearing signals (the UNPROTECTED banner, heartbeat, day P&L, circuit
breaker) are never trimmed or hidden.
CLI board (momentum_cli.py _render_board): drop the static 2-line strategy subtitle (never
changed, redraws every 3s); trim the heartbeat to "Ns"; armed table shows actionable
(armed/fired/long) rows first then the top-4 scanning by rvol, dropping corrupt >1000%-gap /
0-volume rows with a "+N more scanning" footer; remove the redundant ready-signals panel
(restated the armed table) and the day-P&L echo in the trades title; trades float open lots
to the top (never capped off) and cap closed at 6; risk collapses consecutive identical
fires into one "×N" row (the 12 duplicate KEEL lines -> one) and is promoted ABOVE trades.
Dashboard (api/static/dashboard.html): cap the long "runs" — ready-signals to top 15 by
quality (also kills ~80 per-refresh chart fetches), activity feed 20, catalyst feed 20, order
lifecycle shows only actionable orders (raw 1000s incl rejected/cancelled read as an alarm),
watch-states hide the score-gate 'blocked' rows; drop redundant strip cells (Positions dup of
pnl Open/Closed, Gross) and the Account panel's Buying-Power/Synced; demote Account to a
collapsed half-width panel; drop the ~constant ready-signals "Setup" column. JS syntax
validated, no dangling refs, serves 200.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
DB self-heal: the LAN Postgres is reached from WSL via EITHER the 127.0.0.1 netsh
portproxy OR a direct route to 192.168.1.5, and which one works FLIPS across WSL/host
reboots and network-mode changes — the recurring crash-loop landmine. storage/db_pg.py
_connect now tries the configured LAN host and then the alternate, so the loop boots
regardless of the current tunnel state ("connection refused" is instant, so it's cheap).
Dashboard UI: api/static/dashboard.html panels are now collapsible <details> (native,
no JS for the toggle; inner ids preserved so the live rendering JS is untouched). The
live/actionable panels (Armed Triggers, Open Positions, Approval Queue, Ready Signals,
Account) start OPEN and lead the layout; the big/historical ones (Catalyst Engine,
Daily Performance, Trades, Watch States, Activity Feed, Order Lifecycle, All-Time Score)
start COLLAPSED so the page is compact and scannable. Open/closed state persists per
panel in localStorage; an EXPAND/COLLAPSE-ALL header button toggles everything.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The mechanics study found the no-chase entry (buy the VWAP reclaim, not the breakout high)
was the dominant lever. This codifies it as a permanent, pre-registered shadow-validation —
the disciplined close-out of the 4-lever investigation arc.
- strategy/evaluation/vwap_pullback_entry.py: pure find_vwap_pullback_entry — first forward
bar that dips to the CUMULATIVE session VWAP (calculate_vwap, NOT the per-bar minute_bars.vwap
midpoint, which would make 'reclaim' trivially true) and closes back above it; entry = its
close, stop = its low (the tight no-chase structure stop). 4 tests.
- research/labeler.py validate_no_chase_entry + `validate-entry`: simulates it vs the ORB-high
breakout via the live simulate_exit on real forward bars, judged in PERCENT net of a REALISTIC
per-name cost proxy (4% sub-$3 / 2% else — no NBBO in bars). PRE-REGISTERED bar locked: mean
net% must be STRICTLY POSITIVE, 'less negative than ORB' is NOT a pass.
IN-SAMPLE RESULT (v1 gappers, 19 autocorrelated days): the no-chase entry HALVES the loss
(ORB -7.94% -> no-chase -3.61% net, median -4.23%, offered on 75%), robustly (day-jackknife
[-3.83%,-3.48%]) -- confirming not-chasing the spike-top is the strongest mechanic -- but it
still does NOT cross zero under realistic cost: PRE-REGISTERED BAR = FAIL. Consistent with the
whole arc: no mechanizable entry/exit/selection lever beats cost on the full gapper book; the
path to positive is FEWER/cleaner setups (cut the ~70% fader population) COMBINED with this
no-chase entry. validate-entry re-runs forward; promote only if it clears the bar out-of-sample.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The float audit and the labeler both point at the SAME free, available-today selection
signal the bot underuses: gap magnitude + cheap absolute price separate runners from
faders, while the bot's gap%×RVOL rank leans on RVOL (labeler-proven FLAT) and CAPS gap.
- strategy/evaluation/runner_propensity.py: pure, transparent, FROZEN-threshold primitive
-> a coarse runner-propensity tier = gap_tier(0-3) + price_tier(0-2). Gap is weighted
heavier BY CONSTRUCTION because the data says it dominates; premarket gap preferred over
RTH gap; cheap price is a secondary tie-breaker (the two are correlated). 6 unit tests.
- research/labeler.py runner_rank_report + `runner-rank` CLI: measures the lever on the
v1 ORB labeled set. SHADOW/measurement only, not wired into live selection.
MEASURED (v1 gappers, gap>=10%, n=187, 19 autocorrelated days -- small, caveated):
the combined tier ladders monotonically: tier1 2% -> tier3 35% -> tier5 (cheap+big-gap)
48% runner (1.63x base). Gap dominates (>=50% -> 48%, 1.64x); cheap<$2 a real secondary
(1.29x). OPERATIONAL KICKER: the live ~35% TRIGGER_GAP_MAX would EXCLUDE 73 gappers whose
runner rate is 44% vs 20% for the kept <=35% -- the cap discards the best runners (this is
how NXTS at +81% got excluded). Caveat: +1R/+2R separate far less than the +20% runner
label, and gap/price are correlated; a hypothesis to shadow-confirm + adversarially
validate before touching the live cap, not a validated rank.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
GO_LIVE flags `spread_pct` as a not-yet-real field. The reason it was null is
structural: OHLCV bars carry no NBBO, so `minute_bars.spread_pct` can never be
populated from the bars endpoint. The execution-relevant figure is the spread at
the moment a setup is evaluated — a wide book quietly eats R on the thin sub-$20
names this system trades.
This makes that metric real, shadow-first (measure before gating), mirroring how
the catalyst score was introduced:
- `strategy/evaluation/liquidity.py`: pure `compute_spread_pct(bid, ask)` →
(ask-bid)/mid, or None for missing/crossed/non-positive quotes (None never
collapses to a real 0.0). Unit-tested.
- `AlpacaPaperClient.get_latest_quotes`: batched latest-NBBO fetch, mirrors
`get_latest_trades`.
- `Watcher`: optional injected `spread_provider` (same shape as
`catalyst_score_provider`); default None == prior behavior byte-for-byte.
Computed spread is logged onto the criteria_evaluated and signal_ready events.
Observe-only — it does NOT gate ready/blocked.
- `run_live_paper`: TTL-cached spread provider off the client; degrades to None
on any fault, SPREAD_PCT_ENABLED=0 to silence.
The dead `scanner.spread_pct_max` knob is intentionally left for a later
promotion, once the logged data shows spread separates outcomes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019psgRHdtTXz6TCnnqPiqGw
Single-machine equivalent of the Helm package: builds the app from the
Dockerfile, runs PostgreSQL with a persistent volume, and starts the trading
loop with its embedded dashboard on :8010.
- postgres:16 with healthcheck + named volume; app waits on it healthy
- DATABASE_URL overridden to the in-network postgres service (self-contained);
.env is optional (dry mode without ALPACA_* keys)
- journal / eod-replay / nightly-tune under a "jobs" profile, run on demand;
they share the app-data volume so learned_params.json reaches the loop
- Grafana under a "monitoring" profile with a compose-specific datasource
override pointed at the postgres service
- deploy/README.md documents the Compose workflow
Validated with `docker compose config` across all profiles.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JCVMm2AbmZvvxwLTX6UpNR
Document reliability, security, and CI/operational improvements for the
Kubernetes package as an actionable backlog (not yet implemented). Notably
flags the loop-heartbeat liveness gap: /api/health proves only the dashboard
thread is alive, so a hung trading loop would go undetected.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JCVMm2AbmZvvxwLTX6UpNR
Acts on the P2 edge-audit's two findings: (a) ~half of curl fires have a stop tighter
than the spread (median R 1.7% of price, $0.05 on a $2.62 name) so the backtested
+1R/-1R is unfillable; (b) the only subset that survived refutation was price>=$5 AND
mid-tercile impulse, scored as a +2R/runner filter.
- detect_vwap_reclaim: PLACEABILITY GATE (min_r_frac=0.015, min_r_abs=0.02, default ON
for live safety) -- rejects a fire whose R=(entry-pullback_low) sits inside the spread.
signal_values gains r_frac. Tests: a shallow-pullback curl now rejects "stop
unplaceable" and fires only with the gate disabled.
- labeler: validate_vwap_reclaim + `validate-vr` CLI scores the PRE-REGISTERED candidate
(placeable AND price>=$5 AND impulse in [0.15,0.26]) against a LOCKED bar: +2R Wilson-95
lower bound > base 0.349 AND absolute max-upside-60m beats same-symbol-day ORB at
+5%/+10%. Constants + _wilson_lower are committed so the bar can't drift. The shadow
BASE track (vr1) deliberately keeps the gate OFF so report-vr base rates stay complete;
placeability is a filter in validate-vr.
IN-SAMPLE RESULT (decisive): enforcing placeability collapses the audit's n=48 subset to
n=16 -- i.e. most of its apparent +2R lift rode on UNPLACEABLE stops. The remainder FAILS
the locked bar on both criteria: +2R 44% (Wilson LB 23.1% < 34.9% base) and, most
damningly on the artifact-proof axis, it produces a +5%/60m move only 31% of the time vs
ORB's 73% on the same days. Pre-registration did its job: it stopped us chasing an n=48
mirage of sub-spread stops. VERDICT: do not promote; the curl does not beat the bot's
existing ORB even in its best placeable subset. 15/15 unit tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Some sessions have irregular/sub-minute bar timestamps (no exact-duplicate timestamps,
but two bars in the same HH:MM), so keying setup_id on the wall-clock minute let two
fires a cooldown apart collide on the setup_events primary key and aborted the build.
Key on the RTH bar index (unique per session) instead. Tests still 5/5.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
P2 = run detect_vwap_reclaim over history and score it with the SAME no-look-ahead
machinery that scores the live ORB setups, so its base rates are directly comparable
and there's no train/serve skew in the SCORING either (not just the features).
- Extracted research/labeler.py's forward-label logic into a shared _forward_labels()
(max-upside/-drawdown 5/15/60m, reached_1R/2R-before-1R pessimistic, held-VWAP over
the evolving cumulative VWAP, trend-day, failed-breakout, time-to-max). compute_setup
now calls it -- the 3 existing labeler tests still pass, so the extraction is
behavior-preserving.
- compute_vwap_reclaim_setups(): slides the detector over each session's RTH bars and
emits a NON-TRADEABLE setup_event on each RISING EDGE (not-valid->valid, with a bar
cooldown so one curl episode = one setup), carrying the detector's OWN entry/stop/R.
STRICT time split: the detector at bar i sees only bars[:i+1]; labels use only bars
after i. setup_version='vr1', setup_name='vwap_reclaim' -- isolated from the ORB 'v1'.
- build_vwap_reclaim / report_vwap_reclaim + CLI `build-vr` / `report-vr`. report-vr
prints forward outcomes AND a head-to-head vs the ORB baseline on the same symbol-days.
- 2 new tests: a continuation session fires + labels a +1R winner; a one-way fade stays
silent (anti falling-knife). 12/12 green.
PRELIMINARY read (n=35, 4 days -- far too small to trust; full build running): curl-
fires reached +1R 66% / +2R 49%, held-VWAP-15m 86%, avg +8% max-upside in 60m; the ORB
baseline on the same symbol-days reached +1R 67%. Real finding pending the full build.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
P0 anchored the impulse to the day's SINGLE HOD + session-absolute low, so a name
that spiked, topped, then faded read every later bar as "pullback broke the hold"
forever. P1 anchors the leg to a rolling `lookback` window (default 40 bars), so the
detector re-arms on each intraday leg. New regression test proves the property: when
an earlier, HIGHER leg deeply gives back (permanently blocking the whole-session
anchor), the local anchor still fires on a later leg curling off a higher-low base,
targeting that leg's local high rather than the stale day HOD. lookback=None keeps
the legacy anchor for the contrast test. 7/7 unit tests green.
FINDING (ground-truthing the raw bars refuted the P0 premise): MIMI and UPC do NOT
have an RTH VWAP-reclaim setup. Both spiked PRE-MARKET (MIMI 2.93->4.98, UPC
8.54->17.89) and their RTH was a one-way FADE below VWAP (MIMI 4.09->3.15 under VWAP
all day; UPC popped at the open, made lower highs, lost VWAP by 14:17). So 0 RTH
fires is CORRECT, not blindness -- there is no long curl to catch; if the trader
profited it was a pre-market play or a SHORT of the fade, neither of which a
long-only RTH detector should fire on. It fires precisely on the names that DO curl
in RTH (SDOT 6/26 10->22, FCUV 6/25). Same lesson as ignition/PLSM: when the move is
pre-market, the RTH detector should stay silent. (Deliberately did NOT lower
min_run_pct to chase MIMI's sub-10% local legs -- that would be fitting n=2 noise.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The trader's #1 taught setup is a CONTINUATION, not the bot's open-anchored ORB:
a leading gainer that spiked, pulled back (held >=50% of the run), and is curling
back up through VWAP, bought BELOW the HOD. Across 6/25-29 every name he monetized
(MIMI/FCUV/ZDAI/SDOT/UPC) was this one pattern; the bot blocked all on the ORB gate.
P0 = standalone pure detector (detect_vwap_reclaim) with 5 binary gates (impulse /
pullback-held / curl-reclaim / below-HOD / momentum-green), reusing calculate_vwap
& calculate_ema (no train/serve skew). SHADOW-ONLY: not wired into classify_setup.
6 unit tests green.
DECISIVE REAL-DATA FINDING (RTH minute bars): slid across 4 real trader names.
FIRED on the all-day trenders -- SDOT 6/26 (10.22->22.50) 22x, FCUV 6/25 2x -- but
ZERO on MIMI 6/25 and UPC 6/29, both rejected on "pullback broke past the hold
level" with a NEGATIVE held_frac. Root cause is STRUCTURAL: anchoring the impulse
to the day's SINGLE HOD + session-absolute low means a name that spikes->HOD->fades
reads every later bar as "broke the hold" forever, blind to the LOCAL spike->
pullback->curl legs that re-arm intraday. P1 fix = rolling local-swing anchor, not
the day's one HOD. (Deliberately NOT widening hold_frac to force n=4 anecdotes green.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
detect_momentum_ignition (strategy/evaluation/ignition.py): standalone, pure,
binary-rule detector for a vertical squeeze (blue-sky fresh-HOD + price velocity
+ ABSOLUTE volume burst, NOT rvol/float). Stop = last established higher low (per
operator). NOT wired into classify_setup (deliberately). 8 unit tests on
synthetic verticals: fires with the right trigger/stop, and each gate blocks
(not-blue-sky / not-new-HOD / velocity / volume / insufficient-bars / no-ATH).
REALITY CHECK (verify, don't assume): ran it over the REAL PLSM 2026-06-24 bars ->
0 fires, CORRECTLY. PLSM's high was at the 9:35 ET open (it ran +400% PRE-MARKET,
which we don't ingest), and in RTH it was HALTED 228 of ~390 minutes — a
halt-resumption GAP chop (+3.6/-12.5/-17.3/-14.4% jumps), not a smooth vertical.
A consecutive-bar detector structurally cannot see that. So the ignition
opportunity is a PRE-MARKET + halt-resumption phenomenon -> capturing it needs
pre-market bar ingestion + halt-aware gap detection + pre-market/halt trading
(into the FCUV halt-down risk), a far bigger lift than an RTH detector. Documented
front-and-center in docs/IGNITION_SETUP_SCOPE.md as a pre-P1 operator decision.
The P0 detector + tests are a correct foundation regardless.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Scope for a SECOND, shadow-only lane to catch PLSM-class vertical squeezes the
pullback-ORB scorer is structurally blind to (PLSM was evaluated 1895x and
blocked ~30% every time — a vertical gives no pullback). Produced by a read-only
4-agent mapping workflow; no code changed.
Key findings: detect on blue-sky fresh-HOD + price velocity + ABSOLUTE volume
burst (NOT RVOL, which zeros out on thin IPOs; NOT float, which is NULL). Emit a
new ShadowSignalEvent that the execution snapshot (query_ready_signals_snapshot,
filters event_type=signal_ready) can never read -> structurally cannot trade.
Score nightly by reusing the labeler + eod_replay forward-outcome machinery.
Pre-register the promotion bar (runner-vs-trap separation, distinct-runner n)
before shadowing. Phased P0 detector -> P1 dark emission -> P2 nightly scoring ->
P3 validation -> P4 tiny live lever only after separation is proven. Strictly
additive; never touches the ORB book or anti-chase guards.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A small green pop that fades was round-tripping to -1R because the gain
protection only arms at +5% / +1R — SOFI peaked +2.6% (+$281), never reached a
threshold, and gave it all back (stop never moved off the original). Add a
step-trail that ratchets the stop up from entry in step-R increments: at +0.25R
the stop is at breakeven, +0.5R -> +0.25R, +0.75R -> +0.5R, ... It locks a small
pop AND keeps climbing under a runner.
Implemented in the shared _trail_stop (so backtest simulate_exit and live
manage_live agree); risk is recovered from reached_r so it needs no new args;
only ratchets up (max-of-all with the prior-low trail / profit-locks).
TRADING_EXIT_TRAIL_R_STEP (default 0, set 0.25 in .env). Supersedes the interim
standalone +2% breakeven, which reverts to its 0.05 default backstop.
Tests: the ratchet ladder (below-step=no move, +0.25R->BE, +0.5R->+0.25R,
ratchets-up-only) and an end-to-end pop-then-fade that now exits +0.25R instead
of round-tripping to a loss.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root cause of SSPC getting flattened on a mid-session restart (2026-06-24):
flatten_overnight_carries built its today's-buy set from orders with
status=="filled" only. But a marketable buy that PARTIALLY fills (SSPC: 84 of
88) cancels its unfilled remainder, so the order reads status=="canceled" even
though it opened 84 real shares. The carry-detector saw "no filled buy today"
for SSPC, tagged it an overnight carry, and market-closed a position opened ~6
min earlier. (Harmless this time — closed +$20 — but it could cut a runner or
realize a loss.) Marketable orders partial-fill far more than resting limits, so
flipping to marketable exposed this latent bug.
Fix: extract buy_fills_from_orders() (runtime/flatten.py) that counts a buy as a
real today-entry by filled_qty>0 + filled_at, NOT status=="filled" (scans legs
too). flatten_overnight_carries uses it. Pure + unit-tested: partial-fill counts,
full fill counts, zero-fill backout does not, and the partial-fill position is
no longer flagged a carry.
Note: only manifests on a mid-session restart (a clean session's EOD flatten
leaves the book flat, so the open finds no carries). Deploys on next restart.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The last real divergence between the two entry paths was the fill model. Make it
a flag (TRADING_ENTRY_FILL_MODEL), default 'resting' (zero behaviour change):
- resting — rest a buy-limit AT the breakout level (today's behaviour; fills
only on a pullback, misses runners that gap up and never return).
- marketable — limit a hair above the trigger (entry * (1+trigger_slippage_pct)),
fills as price breaks UP through the level (the fast path's
runner-catching behaviour). The limit caps the fill, so a
halt-resume gap-through can't be chased. R is still measured from
the level, not the limit.
Operator chose ship-both-default-resting: flip to 'marketable' to A/B with eyes
on the first session. Boot log shows "fill=<model>".
With gates already unified (steps 1-2: anti-chase, liquidity) the two paths now
run the same gate set and the fill model is selectable — the substantive
unification is done. The remaining single-_place_entry structural merge is
cosmetic (gates are shared) and deliberately not forced through.
Tests: resting rests at the level, marketable lifts above the trigger. 69 green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds the fast path's volume-based share cap to the live auto path: cap shares to
liquidity_max_volume_pct * cum_volume so a thin cheap name can't be oversized
into heavy slippage (the VTAK-style stop that filled far past its level on
2026-06-23). Threads cum_volume through the setup dict + ready-signal snapshot;
applies max_shares in request_approvals_for_ready_signals, gated by
TRADING_UNIFIED_ENTRY. Fail-open: no cum_volume -> no cap. Protective only
(can only shrink a position).
Tests: thin name capped to liq*cum_volume when on, risk-sized when off. 67 path
tests green. Remaining staged (supervised): the full single _place_entry
chokepoint + fill-model unification + halt-status plumbing.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The bot only knew P&L + positions, so "what's on the watchlist?" and "why
isn't it trading?" both fell back to "positions: none" — useless. Now it
retrieves the actual desk state and gives the user direction.
- gather_context: today P&L + recent closed trades (cross-day, robust to the
midnight boundary) + open positions + the live WATCHLIST (states + scores) +
READY SIGNALS (entry/stop/vwap) + a why-blocked tally (risk_rule_triggered
counts) + all-time.
- symbol_context: when a question names a ticker the bot knows, append a focused
dive — its last trade + the setup criteria it passed/failed.
- Help menu (help / /start) lists exactly what it can answer; sent on startup so
the user has direction. Sharper system prompt (concrete, quote real numbers,
say what data exists when it can't answer).
Verified live: "what's on the watchlist?" -> "148 symbols: 97 blocked, 48 late,
3 ready (LAES, NOWL, PBK)"; "why isn't it trading?" -> coherent; "what happened
with VTAK?" -> "stop loss -$659, entry $1.14". Tests: help menu, retrieval
includes watchlist/blocks; 10/10 green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Closes the safety gap the consolidation pass found: the anti-chase / halt
guards (the source-trader blow-up protections) ran ONLY on the fast trigger
path, leaving live auto entries — which is where 100% of fills actually
originate — unprotected.
- Extract _anti_chase_skip(entry_ref, trigger, day_open, halted) from
submit_breakout_now (pure refactor; the fast path is behaviour-identical,
60 trigger/exec tests still green).
- Thread day_open through the setup dict (setup_evaluator) +
query_ready_signals_snapshot so the auto path has the gate input.
- request_approvals_for_ready_signals calls the shared gate behind
TRADING_UNIFIED_ENTRY (env default on). The auto path rests a limit AT the
trigger so it can't chase above it (over_extended is a no-op); the gate that
bites is day-extension — don't enter a name whose breakout level is already
>30% above the session open. Fail-open on missing day_open; halted=False (a
resting limit won't fill mid-halt). Emits rule_type=over_extended_day on skip.
- Boot log shows "anti-chase ON (auto path)".
Tests: _anti_chase_skip logic, parabolic-day blocked when on, allowed when off,
fail-open on missing day_open, normal day allowed. Full path suites green (73).
Note: this is step 1 (the safety win). Full single-chokepoint + fill-model
unification + halt-status-on-auto-path remain staged per docs/CONSOLIDATION_PLAN.md.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Prioritized P0->P2 plan to make the system cleaner + easier to troubleshoot,
produced by a 4-agent read-only mapping pass (no code changed). Key findings:
- Single-Postgres migration is FUNCTIONALLY COMPLETE at the connection layer
(all live code -> storage/db.py -> db_pg -> one DATABASE_URL Postgres; no
live duckdb.connect). Datastore work is cleanup, not architecture.
- The real architectural item is the TWO entry paths. The live (auto) path is
missing five fast-path-only gates: extension, day-extension, halt-guard,
liquidity-sizing, rank — a REAL current gap (47/47 of today's fills were
approved_by="auto"). Root cause: the ready-signal snapshot carries no live
price/day_open/volume/rank. Fix sequenced behind a flag, reversibly.
- Dead-code/cleanup: core/ + telemetry/ packages (zero callers), archive/ tree
+ 232MB zip (gitignored), dead schwab/ integration, an orphaned mem_ test
schema left in prod Postgres.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
C1 — runtime/notifier.py: gated, fire-and-forget push alerts that NEVER raise
into the loop. Decoupled by POLLING the event store (no notifier threaded
through breaker/exit/execution). Pushes on major events only: daily-loss
breaker, catastrophe-stop, naked-stop flatten, any |closed trade| > $300, and a
once-per-day EOD summary. Wired as a 45s scheduler step + EOD summary/failure
hooks in step_eod_flatten. Boot ping on start. Silent no-op when unconfigured.
C2 — telegram_bot.py: standalone (separate process) READ-ONLY Q&A listener,
locked to TELEGRAM_CHAT_ID. Long-polls getUpdates; gathers a Postgres data
snapshot (today P&L, open positions, recent trades, all-time) and asks the
local Ollama model. Falls back to the raw snapshot if Ollama is down/disabled;
refuses any action command (buy/sell/flatten) — questions only.
Tests: disabled no-op, major-event dedup, EOD-once, auth gate, action-refusal,
ollama-down + ollama-disabled fallback. 8/8 green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Wire the one validated selection signal (above-VWAP breakouts reach +1R
~1.5x as often, n=3,110 from the labeler lift report) into the LIVE auto
entry path (request_approvals_for_ready_signals) — NOT the dead fast
submit_breakout_now path where the cap and rank logic got stranded.
- setup_evaluator already computes above_vwap/levels.vwap; surface them in
the setup dict (also an observability win) and pass through
query_ready_signals_snapshot.
- New ExecutionSettings.require_above_vwap (TRADING_REQUIRE_ABOVE_VWAP, env
default on). When a ready signal is below session VWAP: always emit a
rule_type=vwap_below event (shadow measurement); only SKIP the entry when
the flag is on. Fail-open: above_vwap=None (old signals) is never blocked.
- Boot log now prints the entry-gate state, and WARNS if
TRADING_CONCENTRATE_TOP_N>0 (that rank switch only affects the dead fast
path — see the rank-scaling audit).
Tests: gate skips below-VWAP when enforced, shadow-logs+allows when off,
allows above-VWAP, and fail-opens on the missing field.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The per-day fresh-entry cap (TRADING_MAX_FRESH_ENTRIES_PER_DAY) was only
checked in the fast submit_breakout_now path. But live entries flow through
the SLOW auto-approval path (process_auto_approvals -> approve_order), which
had no cap — so the limit was silently dead. Observed 2026-06-23: 44 entries
taken against a configured cap of 5 (all approved_by="auto", zero via the
fast orb_trigger path).
Enforce the cap in approve_order for approved_by="auto", sharing the same
self._fresh_entries counter the fast path uses (so the cap is global across
both paths). Capped orders are REJECTED (reason=daily_entry_cap) so they
leave the pending queue instead of re-skipping every tick. Manual dashboard
approvals are a human override and are never capped.
Two regression tests: auto path binds at the cap + rejects the overflow;
manual approval ignores the cap.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
`python -m research.labeler lift` measures, for each computable signal, the
win/runner rate in its most-favorable bucket vs the base rate (the LIFT), with n
shown so signal is distinguishable from small-sample noise.
First results (n=5,705 setups, full 19 days):
- vs +1R (base 24%): ABOVE-VWAP / distance-from-VWAP is the one real win signal —
35-37% vs 24% = ~1.5x lift on huge n (2,851). gap%/price/data-quality are flat
(~1.0x). Confirms the roadmap's #1 free win: make VWAP a trigger.
- vs +20% runner (base 2.4%): premarket-gap% high-tercile 10.6% = 4.46x (n=426)
is the strongest runner tell; cheap-price low-tercile 6.8% = 2.87x; VWAP/gap
1.4-1.8x. +100% (5-6 events) too sparse to measure -> not a mechanical filter.
Caveat surfaced in the output: autocorrelated days + sparse runners => lift is a
forward-shadow HYPOTHESIS, not a validated win-rate.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
research/labeler.py turns each historical symbol-session into a (features ->
outcome) row so any forecasting signal's LIFT can be measured before it sizes
live. Reuses the LIVE feature fns (opening_range / compute_key_levels /
calculate_time_of_day_rvol / data-quality) so offline == live; STRICT time split
(features <= the ORB-completion decision point, labels strictly after);
idempotent + versioned; populates the previously-empty setup_events /
engineered_features / outcome_labels tables.
`build [--limit N] [--rebuild]` + `report` (base rates + runner rates via one
aggregate query). Core math is a pure compute_setup() — 3 unit tests cover the
time-split (a runner's feature VWAP can't see the post-setup run), the R-outcome,
failed-breakout, trend-day, and the <ORB-bars skip.
First real base rates (n=3,610 setups / 13 days): 24% reach +1R before -1R, 43%
fail outright, +20% runners 2.2%, +50% 0.4%, +100% (NXTS-class) just 0.1%.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The human's edge is concentration (4 trades on 1 name) vs the bot's 12-trade
spray. Added the mechanism: TRADING_CONCENTRATE_TOP_N (only the top-N armed names
by rank may enter; rank-1 full risk, rank-2+ half) + TRADING_MAX_FRESH_ENTRIES_
PER_DAY (hard daily cap). rank threaded from the armed trigger -> submit_breakout_
now -> scales both the % budget and the $500 cap by rank_risk_factor.
LEFT OFF by default (both 0): activating it meaningfully requires the rank key to
be QUALITY, but the armed rank is currently gap*rvol*catalyst. Verification also
found the roadmap's "quality_score" is calculate_data_quality_score = bar
coverage/integrity (a thin-name/liquidity proxy with a time-of-day confound),
correlated only weakly (q>=0.55 = 58% win / breakeven, NOT the 71% claimed; q<0.55
= 35% win / -$2,476). So concentration waits on a sound rank key (goal #2).
pure rank_risk_factor + tests. No behavior change until the env flags are set.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
#3 from the loss diagnosis. Sizing was risk-based (1% of equity) but had no
absolute ceiling, so on the ~$90k paper account every full-budget trade risked
~$900 — the BTBT/AAL/WKSP/LNKS ~-$1k losers (verified: all sized to $900 risk).
Added max_risk_dollars to calculate_position_size: dollar risk used to size is
now min(equity*risk_per_trade_pct, max_risk_dollars), so no trade risks more than
the cap regardless of equity or stop width. ExecutionSettings.max_risk_dollars
(TRADING_MAX_RISK_DOLLARS, 0=off) wired into both entry paths; set to $500 in the
operator .env. PositionSizeResult.risk_amount now reports the ACTUAL sized risk
(after every cap), not the budget.
Verified vs history: the four ~-$1k losers were each sized to $900 risk; under the
$500 cap they'd risk $500 (~halved, ~-$1.8k saved across the four). Small-account
% sizing is unaffected (1% of $300 << $500). Tests: cap binds on big account,
no-op when budget smaller, risk_amount reports actual.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
#1 from the loss diagnosis (the preventive complement to the catastrophe stop).
The entry path already validates geometry and submits a real bracket, so the gap
is POST-FILL: NIVF's stop leg never stayed live, and the exit manager SILENTLY
SKIPPED any position with no usable stop (the `continue` at the old line 160) —
so a naked position got zero management and ran to -23%.
Now the exit manager computes the live stop leg once per position and, if it's
NAKED (no live protective sell-stop), flattens it after enforce_stop_grace_passes
consecutive naked passes (def 2 ~= 24s, a grace so a just-filled bracket's legs
can attach). A position WITH a stop (incl. breakeven-locked, stop>=entry) is
never touched. Long-only via the existing path. TRADING_ENFORCE_STOP_GRACE_PASSES
(0 = off). describe() shows "enforce-stop(Np)".
Net: every held position now either has a live stop or gets flattened within ~24s
— a naked NIVF-class trade can't sit unmanaged to -23% again (and the catastrophe
stop remains the -10% backstop). Tests: naked->flattened after grace; stopped
position never naked-flattened.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Diagnosis (why-strategy-loses-naked-stop): the net loss is ONE naked-stop blowup
— NIVF ran to -23% with no stop that fired (-$3,737 = ~111% of the loss); ex-NIVF
the book is +$380. And the exit manager SKIPS any position without a usable stop
(line ~160), so a naked position gets zero exit management and runs to EOD.
New catastrophe stop runs FIRST and unconditionally on every held position,
before the bars fetch and before that no-stop skip: market-exit (via the robust
cancel-legs-then-close path) when price is >= catastrophe_pct below entry (def
10%, TRADING_CATASTROPHE_STOP_PCT) OR >= catastrophe_risk_mult x the intended 1R
underwater when a stop is known (def 1.5x). Independent of the broker bracket, so
a missing/failed stop can't run to -23% again. Long-only; in-profit and bad-price
are no-ops; 0 disables.
Validated: would have CAUGHT NIVF at -10% (~-$1.6k vs -$3.7k), leaves normal -1R
stops (BTBT -4.2%) alone. Pure `catastrophe_triggered` + manage() wiring +
`_is_active`/describe updated. Tests cover pct arm, risk arm, in-profit, disabled.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The backstop behind the EOD-flatten fix: even with the EOD flatten working, a
carry can survive (loop down at EOD, or the held-qty 403 before it was fixed) and
bleed for days — ATPC sat 4 days over the Juneteenth gap, -$651. This closes any
position carried from a PRIOR session, at the open, before it bleeds.
- runtime/flatten.find_overnight_carries (pure, tested): open positions with no
BUY filled today (ET) = carried overnight. Today's fresh names have a buy
today, so they're never touched.
- ExecutionService.flatten_overnight_carries: detects carries from the broker
(positions + filled buys) and closes them via the robust cancel-legs-then-close
path — WITHOUT halting the session (unlike close_session), so today's trading
continues. Emits an open_catchup_flatten risk event.
- run_live_paper step_open_catchup_flatten: once per session, during RTH, gated by
TRADING_OPEN_CATCHUP_FLATTEN_ENABLED (def 1). Loud log when it flattens a carry.
Dry-run against the live book: correctly targets ONLY ATPC (6/18 carry), leaves
today's DAMD/ECX. 38 runtime/safety tests pass. Needs a loop restart to activate.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Diagnosed the ATPC -$651 drag: on 6/18 (last session before Juneteenth) the EOD
flatten FIRED at 15:55 but every close was rejected — Alpaca 403 "insufficient
qty... held_for_orders" because the bracket stop/TP legs reserved the shares.
That root cause was already fixed by 4520c69 (cancel legs before close; verified
the current self-fetch finds the active leg). What remained were two real
naked-carry holes this closes:
- Early-close half-days NEVER flattened: the window was a fixed [15:55, close),
but half-days close at 13:00 (15:55 > 13:00) — and half-days (Jul 3, Black
Friday, Dec 24) are exactly the pre-holiday sessions. The window is now
computed RELATIVE TO THE ACTUAL CLOSE, so it fires on early closes too.
- No margin / silent failure before a multi-day gap. Before a weekend/holiday
closure the flatten lead widens (TRADING_PRE_CLOSURE_FLATTEN_LEAD_MIN, def 20)
so the book is flat with room to retry, and a flatten that can't fully close
before a multi-day gap now SCREAMS ("‼ NAKED-CARRY RISK (N-day gap ahead)")
instead of failing silently like 6/18 did.
New calendar helpers next_trading_day / days_to_next_session / eod_flatten_status
(pure, unit-tested: normal window, pre-holiday widening, early-close firing,
closed-day skip). Normal full days are unchanged (still 15:55-16:00). Dormant
until the next pre-closure session, so no effect on today's behavior.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
