| name | research-strat |
| description | Research a new trading-strategy idea across the full data-available universe (all instruments x 5m/15m/1h/4h) through the locked horizon + shuffled-null gates and the 8-test battery, dedup vs the live book, and produce a promotion scorecard. Pure research - never touches the locked 21-strategy live book. |
research-strat
Turn a plain-English strategy idea into a graded candidate using the SAME pipeline that
built the deployed 21-strategy book. Search everything, then filter. Output is a
scorecard for a human promotion decision - this skill changes nothing that trades.
Firebreak (NEVER violate)
- NEVER edit
data/tournament/final_universe.json or deploy_book.json (the live book).
- NEVER edit
engine/families.py (the live registry). New ideas go in engine/families_candidates.py.
- All output stays under
data/tournament/candidates/<date>/ (the driver asserts this).
- Promotion into the live book is a HUMAN edit of
final_universe.json at a re-lock point,
never automatic. Present the scorecard and STOP.
Steps
- Translate the idea into a CAUSAL
make_signal(bars, params) -> pd.Series in [-1, 1]
plus a 4-cell param_grid. Causality rule: bar t may use only data up to and
including t; any rolling statistic used as a threshold must be .shift()ed so the
current bar is excluded from its own gate. Add a FamilySpec and register it in
CANDIDATES in engine/families_candidates.py.
- Sweep:
python scripts/research_strat.py --family <name>
Defaults to every data-available instrument x {5m,15m,1h,4h}. Use --symbols /
--freqs to scope, --null-trials N to trade speed for null precision.
- Read
data/tournament/candidates/<date>/<name>.html. Report which (instrument, freq) cells clear the deploy bar and their correlation vs the live book.
- STOP. Present the scorecard. Do not promote, do not touch the live book.
The gates (locked - docs/CONSTRUCTION_SPEC.md + RESEARCH_PLAN.md s4)
- Causality preflight: prefix-stability probe before the sweep. A candidate whose signal peeks
at future bars (shift(-1), centered window, full-sample normalization) is REJECTED outright. CI
(
tests/test_causality.py) covers candidates too.
- Pass A horizon: OOS Sharpe > 0.3, gross > 0, grid-stability >= 70%, cost-share < 60%, maxDD > -25%.
- Pass B shuffled-null: real Sharpe beats >= 90% of signal-permuted nulls (
null_pass_rate <= 0.10).
- 8-test battery (run on the MEDIAN grid cell, not the best - avoids hindsight selection):
T1 CPCV/PBO, T2 regime, T3 grid+sensitivity-regrid, T4/T5 feature+ablation (ML only - fail-closed
if uncomputable; N/A=pass for rules), T6 annual, T7 decay, T8 trade-efficiency.
- Deploy bar: must-pass(T1, T7) AND fail at most ONE applicable test (rules 5/6, ML 7/8)
AND walk-forward Sharpe >= 1.0.
- ADD-READY (the headline verdict): deploy-bar pass AND corr < 0.75 vs every LIVE strat
(
final_universe.json allowlist, not all of combined_daily) AND measured cost available. A cell
that passes the bar but is redundant or cost-unverified shows PASS (...), not ADD-READY.
Notes / known limitations
- Causality is your job in
make_signal. The preflight + CI catch obvious leaks, but write
causal signals: any rolling stat used as a threshold must .shift() so bar t is excluded from
its own gate.
- ML candidate contract: an ML family (name in the ridge/logreg set, or grid carrying
alpha/registry_names/C/l1_ratio) must expose the registry_names feature list so T4/T5
(feature perturbation + ablation) can run. If they cannot be computed, the cell FAILS the bar.
- Cost: missing measured L2 cost falls back to 1x spread (optimistic) and marks the cell
cost-unverified -> not ADD-READY. Run the cost calibration to clear it.
--resume skips cells already in today's JSON (the run persists after every battery cell).- The shuffled-null in
engine/walkforward.py is a full i.i.d. signal permutation (relatively
lenient). A stricter block/circular permutation belongs in that LOCKED path - do not change it
without a re-lock.
- Families needing auxiliary joins (lead_lag, funding_fade, coinbase_premium) need their data
wired before they produce cells; without it those cells skip (recorded in the error list, not
silently dropped).
