dennisthiessen/signal-platform
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs: document post-stop gate reset results

Browse Source
This commit is contained in:
dennisthiessen
2026-07-17 19:58:40 +02:00
parent 5155d00d9e
commit d858475ddb
3 changed files with 223 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+44 -9
View File
@@ -2,13 +2,13 @@
Investing-signal platform for US equities. It runs one strategy, and it is a boring one:
> **A long-only cross-sectional momentum book.** Buy the top quintile by beta-adjusted 12-1 month momentum, tilt toward higher volatility, hold at most 10 names, cut at 1.5× ATR, then trail at 3× ATR for up to 30 trading days.
> **A long-only cross-sectional momentum book.** Buy the top quintile by beta-adjusted 12-1 month momentum, tilt toward higher volatility, hold at most 10 names, cut at 1.5× ATR, then trail at 3× ATR for up to 30 trading days. After an initial-stop exit, re-enter only after the gate has failed and subsequently qualified again.
**Philosophy:** don't predict price — rank it. The edge is *relative* strength across the universe, and the discipline is in the exit: cut losers fast, let winners run until the trail catches them.
**What is NOT the edge — read this before trusting a number on screen.** The composite score, the 5 dimensions, sentiment, fundamentals, and Structural S/R are **display context**, not validated predictors. The Gate Target Ladder is screening machinery that preserves the production setup population; it is not a claim about true market structure. In particular:
- **The headline "target" is not an exit.** It comes from the internal **Gate Target Ladder** and exists only to compute the R:R and reach-probability used by the activation gate. Human-facing chart S/R is a separate model. The live exit reads neither. Across 320 backtested production trades the exit reasons were **144 initial stop, 98 trailing stop, 78 max hold — and 0 targets.** Honoring the target as a take-profit was tested and *halves CAGR* ([research](docs/research/sr-levels-and-exits.md)).
- **The headline "target" is not an exit.** It comes from the internal **Gate Target Ladder** and exists only to compute the R:R and reach-probability used by the activation gate. Human-facing chart S/R is a separate model. The live exit reads neither. Across 472 trades in the current daily gate-reset replay, the exit reasons were **229 initial stop, 147 trailing stop, 96 max hold — and 0 targets.** Honoring the target as a take-profit was tested and *halves CAGR* ([research](docs/research/sr-levels-and-exits.md)).
- **The composite score does not select trades.** Residual momentum does.
Full experiment log — everything tested, kept, and rejected: **[docs/research/](docs/research/README.md)**.
@@ -36,19 +36,29 @@ flowchart TD
BOOK -->|yes| OPEN["OPEN — size at 1% account risk"]
OPEN --> EXIT{"Exit — whichever comes first"}
EXIT --> E1["Initial stop hit<br/>entry − 1.5 × ATR → −1R<br/><b>45% of trades</b>"]
EXIT --> E1["Initial stop hit<br/>entry − 1.5 × ATR → −1R<br/><b>49% of trades</b>"]
EXIT --> E2["Trailing stop hit<br/>highest close − 3 × ATR<br/><i>only binds once price is ~1R up</i><br/><b>31% of trades</b>"]
EXIT --> E3["Max hold reached<br/>30 trading days<br/><b>24% of trades</b>"]
EXIT --> E3["Max hold reached<br/>30 trading days<br/><b>20% of trades</b>"]
EXIT -.->|"NEVER"| E4["Gate Target Ladder target<br/><b>0% of trades</b>"]
E1 --> LOCK["Re-entry locked"]
LOCK --> GF{"Later daily scan<br/>fails the gate?"}
GF -->|no| LOCK
GF -->|yes| GQ{"A subsequent daily scan<br/>qualifies again?"}
GQ -->|no| GQ
GQ -->|yes| RANK
style M fill:#1e3a5f,color:#fff
style OPEN fill:#1e4d2b,color:#fff
style E4 fill:#2a2a2a,color:#888
style E1 fill:#4a1f1f,color:#fff
style E2 fill:#1e4d2b,color:#fff
style LOCK fill:#4a351f,color:#fff
```
**How to read the exit box.** The initial stop is tight (1.5× ATR) and the trail is wide (3× ATR), so the trail sits *below* the initial stop at entry and only takes over once price has advanced roughly 1R. Cut fast when wrong; give room once right. That asymmetry is what produces the right-tailed return profile the strategy depends on — most trades lose a little (win rate ~37.5%), a few win big (best trade +12.9R), and *that is why there is no take-profit*.
**How to read the exit box.** The initial stop is tight (1.5× ATR) and the trail is wide (3× ATR), so the trail sits *below* the initial stop at entry and only takes over once price has advanced roughly 1R. Cut fast when wrong; give room once right. That asymmetry is what produces the right-tailed return profile the strategy depends on — most trades lose a little (win rate 36.2%), a few win big (best trade +12.0R), and *that is why there is no take-profit*.
**What happens after an initial stop.** The stop always closes the trade and realizes its costs. The ticker is then locked until a successful daily full-universe scan first observes it outside the production gate and a later scan observes a fresh qualification. A continuously qualified ticker therefore cannot generate an immediate duplicate entry. Other exit reasons do not start this reset. See the [daily post-stop re-entry study](docs/research/post-stop-reentry.md).
## How It Works
@@ -121,7 +131,7 @@ Once a day (default 07:00). Steps run **in dependency order**, each consuming th
1. **OHLCV** — fetch the latest daily bars for every tracked ticker (Alpaca); new tickers backfill ~5 years.
2. **Sentiment** — fetch sentiment for the names that matter and are stale (> 5 days): top-pick feeders (residual-momentum leaders with a tradeable long setup), the watchlist, and open paper trades, plus a top-N-by-composite discovery net. Runs *before* the scan so the scan sees fresh sentiment.
3. **R:R Scan** — persist clean Structural S/R for charts/alerts, recompute the 5-dimension scores, and build long/short setups from a transient Gate Target Ladder (ATR stops and nominal gate targets) for every ticker. Attach each ticker's residual 12‑1 momentum activation percentile plus the promoted 80/20 production rank.
3. **R:R Scan** — persist clean Structural S/R for charts/alerts, recompute the 5-dimension scores, and build long/short setups from a transient Gate Target Ladder (ATR stops and nominal gate targets) for every ticker. Attach each ticker's residual 12‑1 momentum activation percentile plus the promoted 80/20 production rank. The completed full-universe scan also advances post-stop locks from gate failure to later requalification; failed scans never count as a transition.
4. **Outcome Eval** — resolve setups that hit target/stop or expired (default 30 trading days) and auto-close paper trades per the exit policy (default: 3x ATR trail with a 30-trading-day max hold).
5. **Market Regime** — recompute the regime index (breadth/trend).
6. **Regime Monitor** — separate v2 State/Warning risk thermometer with fixed-basket breadth, VIX, credit, and point-in-time fundamentals; feeds no trades.
@@ -155,6 +165,7 @@ Fundamentals (weekly, early Monday) · Alerts (hourly, Telegram) · Backtest (we
|---|---|---|
| **Residual 12-1 cross-sectional momentum** (the activation gate, long-only) | **Production gate — in-sample edge** | Promoted July 2026 after the portfolio variant beat raw 80 on CAGR, Sharpe and drawdown. Raw 12-1 remains a fallback only when benchmark data is unavailable |
| **3× ATR trailing exit** (+ 1.5× ATR initial stop, 30-day max hold) | **Production exit — best Sharpe of every exit tested** | Beat hold / SMA50 / 20-day-low / technical-40 and both take-profit variants (July 2026) |
| **Post-stop gate reset** | **Production re-entry policy** | The initial stop always closes; the ticker must later fail the daily gate and subsequently qualify again. At the production capacity of 10: Sharpe 1.67 → 1.77, CAGR 45.2% → 48.3%, DD 24.3% → 21.6% versus immediate re-entry. [Full study](docs/research/post-stop-reentry.md) |
| **Structural S/R** | **Human-facing context only — not a gate and not an exit** | Clean, capped zones are persisted for charts and alerts. The scanner deliberately does not read them. |
| **Gate Target Ladder** | **Gate input only — not market structure and not an exit** | Volume-free range grid + pivots preserves the useful legacy screening behavior exactly: 1,086/1,086 qualified setups retained and identical Sharpe 2.03 / CAGR 50.0% / DD 21.4% / 321 trades. The exit never reads its target. [Full write-up](docs/research/sr-levels-and-exits.md#explicit-gate-target-ladder) |
| Composite score + 5 dimensions | **Display/ranking only** | Sub-scores are hand-built heuristics; none has a measured IC. Note: the "momentum" *dimension* is 5/20-day ROC — NOT the validated 12-1 factor (that lives in `momentum_service`) |
@@ -167,11 +178,25 @@ Fundamentals (weekly, early Monday) · Alerts (hourly, Telegram) · Backtest (we
Caveats on the momentum result: in-sample, roughly one market regime, costs/slippage approximated at 0.1% per side, and residual momentum still needs SPY benchmark history to compute. The **out-of-sample proof is the forward paper-trade record**: Signals → Track Record compares live qualified expectancy against the backtest.
### Current production baseline
### Daily post-stop re-entry decision (2026-07-17)
Use this as a regression guardrail for future strategy changes, not as a return promise. Backtest run: local production SQLite snapshot, 506 tickers, weekly cadence, 30-trading-day horizon, 2022-06-28 → 2026-07-02, 0.1% per-side costs, price-only SPY benchmark. Numbers below are the 2026-07-11 run (`reports/backtest-20260711-prod-baseline.json`) — measured *after* the primary-target probability floor shipped, which pruned lottery-target setups (1,428 → 1,089 qualified) and lifted Sharpe on all three promotion contenders.
The production policy is **normal gate reset**, evaluated with daily setup opportunities and live-like full-universe ranking. An initial stop always closes. Re-entry unlocks only after a later successful daily scan observes the ticker failing the gate and a subsequent scan observes it qualifying again. The study replayed 1,011,248 point-in-time candidate observations across 505 tickers from 2022-06-24 through 2026-07-02, with the production GTL gate, 80/20 rank, exit, fees, sizing, and 10-position capacity.
| Item | Current baseline |
| Re-entry policy | Total return | CAGR | Max DD | Sharpe | Trades |
|---|---:|---:|---:|---:|---:|
| Immediate | 348.4% | 45.2% | -24.3% | 1.67 | 489 |
| **Gate reset (production)** | **388.1%** | **48.3%** | **-21.6%** | **1.77** | **472** |
| Fixed five-session cooldown | 250.8% | 36.6% | -22.2% | 1.47 | 473 |
In the disjoint 2025+ book, gate reset also beat immediate re-entry (Sharpe 1.66 vs 1.55; CAGR 41.8% vs 39.3%) and the fixed five-session rule (Sharpe 1.43; CAGR 32.7%). Its lead over both survived costs of 0.2% and 0.3% per side. The result is capacity-specific: cooldown 5 won at capacity 5, while immediate had slightly higher return and Sharpe at capacity 15. Production uses capacity 10, so that is the portfolio for which this decision is valid.
`gate_reset` and a simple `next_session` block happened to produce the same executed live-universe portfolio in this sample. Their rules are still different: this establishes that same-day re-entry was harmful here, but does not isolate a separate historical return premium from the reset condition. Gate reset was promoted because it represents a genuinely new signal episode and did not sacrifice results in the production book. Full definitions, all nine policy arms, cost/capacity sensitivity, and legacy-rank results are in [docs/research/post-stop-reentry.md](docs/research/post-stop-reentry.md); source report: [`reports/daily_reentry_matrix.json`](reports/daily_reentry_matrix.json).
### Historical weekly production baseline (pre gate-reset)
Use this as the historical ranking/exit regression guardrail, not as a return promise or the current re-entry-policy result. This run predates the post-stop gate reset and uses weekly entry replay, so its portfolio headline is not directly comparable with the daily matrix above. Backtest run: local production SQLite snapshot, 506 tickers, weekly cadence, 30-trading-day horizon, 2022-06-28 → 2026-07-02, 0.1% per-side costs, price-only SPY benchmark. Numbers below are the 2026-07-11 run (`reports/backtest-20260711-prod-baseline.json`) — measured *after* the primary-target probability floor shipped, which pruned lottery-target setups (1,428 → 1,089 qualified) and lifted Sharpe on all three promotion contenders.
| Item | Historical weekly baseline |
|---|---|
| Strategy version | `residual_highvol_80_20_atr_trail3_v1` |
| Production gate | Long-only, residual 12-1 momentum percentile >= 80, headline gate-target R:R >= 2.0 (live `activation_min_rr`; the code default is 1.2), primary-target reach-probability >= 20%, NEUTRAL excluded, confidence floor off (0) |
@@ -218,6 +243,7 @@ A systematic single-variable sweep (offline prod snapshot, production gate/rank/
| Momentum lookback: 6-1, 3-1, 12-7 (Novy-Marx), composites | **Keep residual 12-1** | 6-1/3-1 rank-IC ≈ 0; 12-7 IC 0.045 / t 1.58 — weaker than residual 12-1 (0.055 / t 1.98) |
| Selection cutoff {70, 75, 85, 90} × book size {10, 15, 20} | **Keep 80 × 10** | Monotonically worse in both directions from 80; the 10-slot cap never binds (<10 concurrent) |
| Position sizing: equal-weight, inverse-vol, risk-% sweep | **Keep 1% fixed-fractional** | See the inverse-vol warning below |
| Post-stop re-entry: immediate, fixed 2–5 sessions, gate resets, confirmation filters | **Keep normal gate reset for the 10-position production book** | Sharpe 1.77 vs 1.67 immediate and 1.47 cooldown 5; rerun before changing portfolio capacity |
| FIP path-smoothness as an in-book tie-breaker/filter | **Reject** (but see the lead below) | Non-monotonic across FIP quintiles within the qualified set; either half of a median split underperforms the full book — thinning the entry stream costs more compounding than the tilt returns |
Two findings future sessions must not re-litigate:
@@ -450,6 +476,14 @@ python scripts/run_backtest_snapshot.py backtest_snapshots/prod.sqlite --workers
.venv\Scripts\python.exe scripts\run_backtest_snapshot.py backtest_snapshots\prod.sqlite --workers 6 --allow-spawn
```
Weekly remains the resource-safe default. Add `--cadence daily` for live-like daily entry opportunities; this performs roughly five times as many setup evaluations. To generate the complete weekly/daily × immediate/gate-reset comparison in one invocation, use:
```bash
python scripts/run_backtest_cadence_comparison.py backtest_snapshots/prod.sqlite --workers 7
```
On Windows, add `--allow-spawn`. The comparison runner writes the two full cadence reports plus one compact four-arm report. For the larger nine-policy daily research matrix used in the post-stop decision, see `scripts/run_daily_reentry_matrix.py` and the [research record](docs/research/post-stop-reentry.md).
On an 8-thread machine, `--workers 6` is a good starting point: it leaves a
couple of threads for Windows, the shell, and browser/UI work while still using
most of the CPU.
@@ -490,6 +524,7 @@ matching decision. Every change still goes through the factor harness first (see
| `gate_ablation` | Net expectancy with each floor removed | Drop a floor only if removing it doesn't hurt net expectancy |
| `time_exit_sweep` | Net avg R / net R-per-day by hold length | Whether a fixed time exit beats the promoted ATR trail |
| `portfolio_monitor`, `portfolio_sim`, `strategy_variants` | CAGR, Sharpe, max drawdown, per-year returns | Promote a strategy only if it beats the current baseline on CAGR/Sharpe/DD |
| `production_cadence_comparison` | Immediate vs production gate reset at the selected weekly or daily cadence | Isolates the re-entry rule while keeping gate, rank, exit, fees, sizing, and capacity fixed |
| `signal_eval` | Mean IC, t-stat, IC>0 %, `reliable` | Iron rule: wire a new factor in only if \|IC\| ≳ 0.03 with a consistent sign and `reliable: true` |
| `holdout` (opt-in) | Train vs test books, split by entry date | **The only honest OOS read.** Set `BACKTEST_HOLDOUT_SPLIT=YYYY-MM-DD` |
| `recommendation`, `research_recommendation` | The report's own headline read | A starting point, not a substitute for the sections above |