min_rr = 2.0 was hand-set in Admin (2026-06-24) and never swept — the gate
ablation only tested the floor on-vs-off, never its level. It was the last
un-swept knob in the live gate.
Swept against portfolio Sharpe under the real exit, with a parity self-check
(reproduces_production_gate: the row at the live floor must rebuild production's
exact 1,089-setup qualified set — it does).
min_rr qualified in-sample Sh/CAGR OOS Sh/CAGR (entries >= 2024-07)
0.0 6636 1.98 / 58.5% 2.02 / 66.2%
1.2 3897 1.34 / 33.9% 1.12 / 28.8%
1.5 3127 1.20 / 29.6% 1.12 / 28.8%
1.75 1974 1.64 / 44.5% 1.15 / 27.4%
2.0 (live) 1089 2.04 / 50.4% 2.78 / 73.3%
2.25 577 1.64 / 31.8% 1.71 / 31.9%
2.5 286 1.67 / 29.0% 0.68 / 8.7%
KEEP 2.0. It is the optimum in both windows, and a peak that reproduces in data
it was never fitted to is real evidence. But treat it as fragile: unlike the ATR
trail (a plateau), this is a spike with a trough beside it — +/-0.25 costs ~0.4
Sharpe in-sample and ~1.6 out-of-sample — and the curve is bimodal (floor-off is
good, 1.2-1.75 is bad, 2.0 is good). The hand-set value landed on the peak by
luck, not by tuning. Do not nudge it.
Worth knowing: turning the floor OFF entirely is the second-best row in both
windows, with substantially higher CAGR (58.5% / 66.2%) and more trades. If CAGR
ever outranks Sharpe here, "no R:R floor" is a live option — and it would sever
the gate's last dependency on the weak S/R detector.
Also fixes a metric artifact in the holdout harness. The train book's equity curve
ran to the end of the data while its entries stopped at the split, so it sat in
flat cash for two years and deflated its own CAGR/Sharpe (reported 0.95 / 14.6%;
actually 1.31 / 29.6%). _simulate_portfolio now truncates the calendar to
hold_days after the last entry when end_date is set — it only triggers on the
holdout train window, so no other number moves. The clear-air OOS verdict is
unaffected: it rests on the test row, whose entries and curve both start at the
split and were always clean. Both holdout reports regenerated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Investigated whether our support/resistance detection follows best practice
and whether we actually use it that way. Three findings, all backed by runs
against the prod snapshot and written up in docs/research/sr-levels-and-exits.md:
- The S/R target must NOT become an exit. Honoring it as a take-profit on top
of the 3x ATR trail drops Sharpe 2.04 -> 1.47 and halves CAGR. Win rate rises
(37.5% -> 40.0%), which is the tell: it truncates the right tail where
momentum's edge lives.
- The clear-air fallback (synthesize a 3xATR target where no resistance exists,
so 52-week-high breakouts stop being vetoed) looked strictly better in-sample
(Sharpe 2.04 -> 2.07, CAGR 50.4% -> 62.3%, DD 21.4% -> 20.1%) but FAILED a
real out-of-sample holdout: on entries after 2024-07-01 it is worse on Sharpe
(2.78 -> 2.45) and Calmar, better only on raw CAGR. Not shipped.
- The detector itself is weak vs best practice (POC/VAH/VAL computed then
discarded, HVN = any above-mean bin, 1.48x volume double-counting, "touch"
counts pass-throughs, no round numbers), but its only causal path to P&L is
the entry gate. Fix it for the displayed levels, not for returns.
Method note: nested lookback windows are NOT out-of-sample. The in-sample result
was clean, large, and consistent across five windows, and still did not survive
a proper entry-date split.
All research paths are off by default and the default report is unchanged:
BACKTEST_RESEARCH_EXITS=1 take-profit exit rows
BACKTEST_ATR_TARGET_FALLBACK=k synthetic k*ATR target when S/R offers none
BACKTEST_FALLBACK_CLEAR_AIR_ONLY=1 restrict that to genuinely clear air
BACKTEST_HOLDOUT_SPLIT=YYYY-MM-DD train/test split by entry date
Also fixes two reproducibility holes found while reconciling our local baseline
against the live report:
- create_backtest_snapshot.py now copies paper_% settings. The production
monitor row replays the runtime exit policy via get_exit_policy(); without
those keys a snapshot silently falls back to code defaults, so a live-tuned
exit would never be reflected.
- Migration 020 drops activation_min_expected_value and
activation_min_target_probability. Both are orphans of the June EV-gate
redesign, read by no code path, but prod carries min_target_probability = 50.0
which implies a probability floor that is not enforced (the real floor is the
20% constant in qualification.py).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three follow-ups to the gate probability floor (8f41143):
- Signals table shows the starred primary target (shared primaryTarget
helper) instead of an independently computed max-probability best,
so Overview, Signals and ticker details agree by construction.
- Targets pinned at the 3% probability clamp floor collapse to the
nearest one (enhance_trade_setup + backtest candidates in parity):
floor-pinned levels are indistinguishable to the model, so farther
ones were duplicate 3% rows inviting lottery headlines.
- get_trade_setups only returns setups re-emitted within
LIVE_SETUP_MAX_AGE_DAYS (3): an older latest row means the daily
scan no longer confirms the setup, and such rows otherwise surface
forever on Overview/Signals/ticker/alerts. History endpoints keep
full history.
Backtest on the Jul-3 snapshot is metric-identical to the gate-floor
run on all qualified stats (1089 qualified, Sharpe 2.02, CAGR +49.6%,
DD -15.8%): the prune only removes noise the gate already rejected.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The portfolio monitor's Production row now replays the live qualification
flag and the Admin exit policy (mode/ATR multiplier/hold days) instead of a
frozen research-variant gate, so Admin tuning is reflected in the next run.
Single-source the 80/20 strategy_rank weights in momentum_service and pin
every dual-defined constant with a parity test. Behavior-preserving today:
the production sim reproduces the README baseline exactly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Fixes ruff E741 in the lookbacks comprehension of _portfolio_monitor,
which failed the CI lint step and blocked the deploy.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Backtest report now includes research-only hold-to-horizon portfolio variants comparing raw vs residual 12-1 momentum, cutoff 80 vs 90, max 10 vs 15 positions, and SPY-200 risk scaling. A dynamic research recommendation panel flags residual momentum, cutoff 90, or regime scaling only when transparent promotion rules pass.
Adds signal_context_snapshots with migration 016 and captures one point-in-time context row per newly generated TradeSetup: setup fields, composite/dimensions, latest sentiment, latest fundamentals, and strategy_version=momentum_12_1_rr_time_v1. This is forward-only; no historical sentiment/fundamental backfill is attempted.
No live gate, paper-trade exit, or production ranking behavior changes.
Verification: 458 backend tests pass, ruff check app/ clean, frontend npm run build clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds a research-only 12-1 residual momentum signal to the cross-sectional signal-evaluation harness. The signal estimates benchmark beta over the 12-1 formation window and ranks cumulative stock return minus beta-adjusted benchmark return; it only appears when benchmark closes are available.
No production qualification behavior changes. The Backtest signal table labels the new row as 12-1 residual momentum. Tests cover benchmark-gated emission and beta removal while keeping stock-specific drift.
Verification: 453 backend tests pass, ruff check app/ clean, frontend npm run build clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The robustness warning was computed on the target-model distribution
while the same panel recommends the hold exit — internally inconsistent.
_robustness_stats (median, profit factor, ex-top-5% expectancy) is now
shared by _bucket_stats and _time_exit_bucket, the time-exit table shows
Median Net R and Ex-Top-5% per hold length, and _build_recommendation
reads the trimmed expectancy from the recommended exit's bucket (falling
back to the target model when no hold is recommended).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Robustness (answers 'is the edge just outliers?'):
- _bucket_stats gains median_net_r, profit_factor, and net_avg_r_ex_top5
(expectancy with the top 5% of winners removed); shown as stat tiles.
- Portfolio sim gains per-calendar-year returns, shown in the sim table.
Dynamic recommendation ('What this backtest recommends' panel):
- _build_recommendation derives advice from the report's own numbers on
every run — exit policy (target vs best hold, with sim CAGRs), which
gate floors earn their keep (ablation Hold column), best momentum
cutoff, book-vs-SPY verdict, and an outlier-dependence warning when
the trimmed expectancy goes non-positive.
Retired (conclusions reached, tables removed from report + UI):
- Take-profit sweep (no interior optimum — fixed TP is the wrong tool
for momentum), trailing sweep (converged to the hold-to-horizon exit),
probability calibration (model is display-only by decision).
- _tp_primitives slimmed to _risk_and_stop_day; trailing machinery gone.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Per-trade additions to the report:
- Gap-through-stop fills: stops now fill at the worse of the stop or the
bar's open across every exit model (target, TP, trailing, time), so a
loss can exceed -1R; targets never fill better than their level.
- best_r / worst_r, avg holding days, and net R per day of capital
deployed on the summary buckets and the time-exit sweep.
Portfolio simulation (the stats a per-setup replay cannot give):
- One capital-constrained book over the qualified setups: 10k start, max
10 concurrent positions (one per ticker, best momentum first), 1%
fixed-fractional risk with a 20% no-leverage notional cap, entries at
the detection close, 0.1%/side costs, daily mark-to-market.
- Two exit policies compared: S/R target race vs hold-to-horizon.
- Equity-curve stats: final equity, total return, CAGR, max drawdown,
annualized daily Sharpe, win rate, avg P&L, best/worst trade, avg
hold, entries skipped on a full book, and SPY price return over the
same window (benchmark history refreshed to cover the replay span).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The ablation judged floors under the target/stop model, but the exit
sweeps point at replacing that exit with a fixed hold — under which the
R:R floor's rationale (bigger payoff at the target) may not apply. Each
ablation row now also carries hold_avg_r / hold_net_avg_r / hold_total_r
(30d hold, initial stop only), so the Phase 3 gate decision can be read
under the exit policy that would actually be used.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
_window_setups computed them but _replay_ticker dropped them, so the
ablation's NEUTRAL/tightener checks saw None for every candidate and the
'without confidence floor' / 'without R:R floor' rows collapsed to 0
setups (impossible — removing a floor can only add setups). Regression
test now goes through the real _replay_ticker path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Phase 1 of the strategy-measurement plan — report-only, no production
trading behavior changes:
- Cost haircut: every bucket/sweep now reports net_avg_r/net_total_r
alongside gross (COST_PER_SIDE=0.1% of notional, converted to R via
each setup's stop distance); params carry cost_per_side_pct.
- Gate ablation table: re-qualifies candidates at the current momentum
cutoff with one floor removed per row (confidence / R:R / NEUTRAL /
momentum-only) to show which floors earn their keep.
- Time-based exit sweep: hold 5/10/21/30 days with the initial ATR stop,
exit at the day-N close — the classic momentum implementation, to
disambiguate the wide-trailing result.
- TP sweep extended to +40/+50%, trailing to 25/30% so the optima are
interior instead of starred at the sweep edge.
- BacktestPanel: Net Avg R columns everywhere, gate-ablation and
time-exit tables, stars now mark best net avg R; stale cached reports
still render (all new fields optional/guarded).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Third exit model alongside target-vs-stop and the fixed take-profit. The TP sweep
showed the edge lives in the fat tail (avg R keeps rising as you let winners run),
but a fixed wide target is win-rate-brutal and gives everything back on a reversal.
A trailing stop harvests the tail while protecting gains.
Per setup the replay computes the realized R for several trail widths (3/5/7/10/
15/20%) in a single conservative pass — stop ratchets up via max(initial_stop,
peak*(1-trail)), exit on the pullback or at the horizon close, R vs the initial
risk. Aggregated into a trailing sweep (win rate = share closed in profit, avg R,
total R) over the qualified set and shown as a new table in the Backtest panel.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Avg R was still rising at the previous top level (+15%), so the optimum was off
the table. Extend TP_LEVELS to 20/25/30% to reveal where letting winners run
stops paying (it plateaus toward "just hold to the horizon close").
Also clarify in the panel that the take-profit model deliberately does NOT use
the setup's S/R target — it's a standalone fixed-% exit; exiting at the target is
the target-vs-stop model above. The two are complementary ends, not in conflict.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The target-vs-stop model counts a near-miss of a far S/R target as a full loss
and ignores the partial gains you actually bank — so it measures a different
strategy than "scalp the early pop, take +8%". Add a realistic take-profit exit
model next to it (original untouched).
Per setup the replay now also records risk%, whether the stop was hit, the
favourable excursion reachable before the stop (MFE), and the horizon-close move.
From those a fixed-take-profit sweep (4/6/8/10/12/15%) is scored in R: bank +X%
if reached before the stop, else -1R, else the horizon close. Hit rate = how
often +X% was banked (the MFE CDF), so you can pick the EV-optimal TP without
top-ticking fantasy. Shown as a new table in the Backtest panel; the IC,
calibration and momentum sweep are unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Behavior-preserving cleanup (345 tests pass, ruff clean):
- scheduler: replace 62 inline logger.x(json.dumps({...})) calls with a
_log_event helper, and collapse 11 identical _job_runtime dicts into an
_idle_runtime() factory over _JOB_NAMES.
- settings: add app/services/settings_store.py (get_setting/get_value/get_map/
upsert_setting) and route ~13 hand-rolled SystemSetting queries + two
identical _settings_map helpers through it.
- scoring.get_rankings: collapse the per-ticker N+1 (3-4 queries + a commit each)
into 2 bulk reads + a single conditional commit; drop the redundant re-fetch.
Lazy recompute-on-read is preserved. Adds first tests for get_rankings.
Net ~ -245 lines across the touched modules.
Part 1 — long-only. The momentum edge is long top-momentum; the gate was
qualifying shorts on high-momentum names (fighting the trend), which showed as
the -0.13R Short(qual.) drag. While the gate is active, shorts no longer qualify
(backend qualification, backtest _momentum_qualifies, and the frontend mirror).
Part 2 — production wiring. Live setups now carry a real momentum rank, so the
dashboard, the Track Record's qualified stats, and outcome evaluation all gate on
the same value instead of deferring to floors:
- new momentum_service.compute_momentum_percentiles: 12-1 momentum per ticker,
ranked across the universe into a {symbol: percentile} map.
- the daily R:R scan ranks the universe up front and stores each setup's
percentile (new trade_setups.momentum_percentile column, migration 010).
- enhance_trade_setup mutates the same row, so the percentile is preserved;
_trade_setup_to_dict + TradeSetupResponse expose it to the API.
Until a fresh scan runs, pre-existing setups have a null percentile and the gate
falls back to floors for them (longs) / excludes them (shorts) — they fill in on
the next scan. 341 backend tests pass; frontend build clean.
Needs the alembic upgrade (migration 010) on deploy.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The replay was CPU-bound and single-core: the earlier asyncio.to_thread offload
kept the API responsive but, because of the GIL, ran on one core. Per-ticker
replay is independent, so fan it out across worker processes (which sidestep the
GIL) for real multi-core speedup.
- New `settings.backtest_workers` (default 4), capped to cpu_count-1 so a core
stays free for the web server.
- Uses a `forkserver` context (workers forked from a clean single-threaded
server — avoids the fork-with-threads deadlock); falls back to `fork`. On
spawn-only platforms (Windows) and for 1-ticker runs it uses the thread path,
so dev/tests are unaffected.
- Worker takes primitive column arrays (cheap to pickle), rebuilds bars, and
returns (candidates, plain-dict signal series) — both picklable across the
process boundary. Bars are still fetched in the event loop (ORM-safe).
- Pool creation is guarded: if the pool can't start, the job falls back to the
sequential thread path instead of failing.
334 backend tests pass (parallel path is POSIX/server-only, so it's covered by
construction + the picklability/worker-count tests; the thread fallback is
exercised by the run_backtest smoke test).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The 5-year backtest confirmed the EV gate adds negative value (high threshold =
worst expectancy) and that 12-1 month momentum is the one price signal with a
plausible, right-signed cross-sectional IC (~0.05). So "qualified" now means:
clears the R:R + confidence floors AND the ticker ranks in the top
`min_momentum_percentile` of the universe by 12-1 momentum that week.
- qualification.py: drop expected_value_r / the EV gate; add a momentum-percentile
gate (duck-typed `momentum_percentile`, only enforced when attached + threshold
set, else defers to floors). Mirrored in frontend qualification.ts.
- activation config/schema: min_expected_value -> min_momentum_percentile
(default 80 = top quintile). ActivationSettings, DashboardPage (ranks/【shows】
momentum instead of EV), and the BacktestPanel sweep follow.
- backtest: rank each ISO week's universe by 12-1 momentum, assign a percentile,
and qualify the top slice; the sweep now sweeps the percentile cutoff.
Also offload the backtest's per-ticker compute to a worker thread so the heavy
~5y run no longer blocks the API event loop (the "backend offline" flicker).
Production setups don't carry momentum_percentile yet — wiring the scanner to
attach it (a universe momentum-rank step) is the next step; until then the live
gate defers to floors while the backtest measures the momentum selection. 330
backend tests pass; frontend build clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two changes so the cross-sectional signal results can actually be trusted.
(a) History depth — the binding constraint. Ingestion defaulted to 365 days, so
long-lookback factors (12-month momentum, 52-week high) were only computable on a
handful of weeks at the tail, and every IC reflected a single market regime.
- New `settings.ohlcv_history_days` (default 1825 ≈ 5y); new tickers backfill this
far instead of 1 year.
- New manual "data_backfill" job (Admin → Jobs) re-fetches the full window for
every ticker, ignoring incremental resume — run once to deepen existing
1-year histories. Idempotent (upsert); resumes after rate limits.
(b) Factor-IC honesty. The IC was averaged over weekly rebalances whose 30-day
forward windows overlap, inflating the t-stat ~sqrt(6)x.
- IC now measured on NON-OVERLAPPING windows (weeks thinned to ~HORIZON apart).
- Each signal carries a `reliable` flag (>= 12 independent windows); BacktestPanel
greys out and de-stars thin signals so a lucky 9-week IC of 0.3 can't masquerade
as an edge.
332 backend tests pass; frontend build clean. No migration (config + job + an
added JSON field on the cached backtest report).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The per-setup hit-rate report can't tell whether a signal predicts returns —
only how a target/stop structure built on one performs. This adds a
cross-sectional factor-IC pass: each week the universe is ranked by a price-only
signal and graded by its rank correlation (Spearman IC) and top-minus-bottom-
quintile spread against the forward 30-day return.
Candidate signals (point-in-time from price; sentiment/fundamentals have no
history in the replay): 12-1/6-1/3-1 month momentum, 1-month reversal,
price-vs-200d SMA, proximity to the 52-week high (George/Hwang), and 126-day
realized volatility (low-vol anomaly).
Reuses the existing per-ticker replay loop (no new data, no second DB pass);
results land in the cached backtest_report as `signal_eval` and render as a
"Signal edge" table in BacktestPanel beside the calibration curve.
330 backend tests pass (10 new in test_signal_eval); frontend build clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Diagnosing "no qualified signals for 5 days": setups were generated but none
qualified. The gate required BOTH a high min_rr (2.0) AND a high
min_target_probability (60), which became contradictory after the Jun-15
probability recalibration — probability already embeds R:R via the 1/(rr+1) ruin
term, so high-R:R targets are inherently low-probability and nothing cleared both.
Gate is now expected value (R): p*rr - (1-p) from the primary target's
probability. R:R and confidence stay as floors; high-conviction / exclude-conflicts
/ min-target-probability become optional tighteners (default off). Defaults:
min_expected_value=0.15, min_rr=1.2, min_confidence=55. EV is only enforced when
computable. Migration 009 clears stored activation_* rows so the new defaults
apply. Backtest sweeps min_expected_value instead of target probability.
Scheduling: pipelines are now cron-configurable in Admin -> Jobs. daily_pipeline
(full, default 0 7 * * *) plus a new light intraday_pipeline (OHLCV + outcome eval,
default hourly US session) that keeps prices/live-R:R current without setup churn.
Fundamentals on its own early weekly cron. Timezone configurable (default
Europe/Berlin). Moving interval->CronTrigger also fixes the restart-deferral bug
where an interval job's countdown resets on every process restart.
319 backend unit tests pass; frontend tsc clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Re-applies the activation gate at several min_target_probability thresholds
(60→30, other conditions fixed) over the already-replayed candidates, so the
trade-off between how many setups qualify and their expectancy is visible in one
table — the cheap "optimize" half of Phase 2. Candidates now carry meets_core +
best_prob so the sweep needs no re-replay. New sweep table in BacktestPanel with
the current threshold starred.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replays the price-derived engine over stored OHLCV: at each weekly as-of date,
rebuild the setup from bars <= D (no lookahead) and walk the actual forward bars
for the realized outcome. Reports realized hit-rate/expectancy of qualified
setups (and all setups, by direction) plus a probability calibration curve
(predicted target prob vs realized hit rate).
Reuses pure functions throughout; extracted compute_technical_from_arrays /
compute_momentum_from_closes from scoring_service so live and backtest stay in
sync. Runs as a weekly/triggerable 'backtest' job caching the report in a
SystemSetting; GET /backtest/report serves it. Sentiment/fundamentals held
neutral (no point-in-time history) — calibrates the price/S-R/probability machinery.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
