🧠 Freebuff Session Log — Money-Ready Methods Work

Oct 19, 2026 · Buffy (Freebuff / Codebuff) · independent lane: statistical proof layer · page: freebuff-2026-10-19-money-ready-methods.html

1. What this session shipped (changes)

✅ Task 10 — `alpha_engine/benchmark_closed_ledger.py` (new)

The plan named this module; it didn't exist. It benchmarks every closed-pick ledger against (a) its same-window benchmark via a paired bootstrap (publish gate: ≥95% of resamples must beat the benchmark) and (b) a random-entry Monte-Carlo (10k shuffled entries, same hold lengths — percentile ~1.0 = timing edge, ~0.5 = coin flip). Self-contained stdlib stats core; explicit benchmark map (EQUITY/ETF→SPY, CRYPTO→BTC-USD); WR hidden below n=20; fail-soft with reasons, never exceptions.

Verified: tests/test_benchmark_closed_ledger.py 33/33 pass (exit 0) · py_compile clean · real-data run exit 0 → reports/benchmark_ledger_proof_2026-10-19.json (170 ledgers, 67 with full proof, 6 of 170 clear the paired benchmark gate — 4 within the full-proof subset). After external review the random-entry MC was fixed (same-asset, window-matched, direction-aware nulls; the old version sampled BTC/SPY paths and pinned a PF-1.03 ledger at percentile 1.0 — post-fix that ledger scores 0.57, a coin flip; a CI coin-flip gate test guards the fix). Live read-only panel: /audit/benchmark_proof.html. Doc: updates/2026-10-19-task10-benchmark-closed-ledger.md.

🔍 Method review — "what actually works" (report)

reports/METHOD_REVIEW_2026-10-19-what-actually-works.md — all 6 candidate methods audited against repo evidence, every figure machine-verified (checker exit 0: 5 backtest files verbatim, 9 forward cells recomputed, funding/PEAD data files, kill gates). Headline: TSMOM is the only method with Tier-2 WF backtest + PF>1 forward (ETF sleeve n=35, PF 1.76); PEAD is walk-forward-verified but has zero production trades (shadow-only, never wired); funding carry regime inverted (mean −189.7%/yr on the frozen June snapshot); QMJ/value forward PF 0.69–0.88; pairs 1 win of 8; vol premium has no-cost simulation only.

2. Achievements (verified outcomes, not intentions)

Task 10 real-data findings

Ledger	Class	n	PF	Excess vs bench	frac+	MC pct	Read
mega_mutation_macd_rsi_m048	CRYPTO	57	5.11	+5.88%	1.00	0.99	clears both gates — paper-pilot candidate only (class NOT_READY)
forex_rsi2_mean_reversion	FOREX	21	3.88	+0.57%	1.00	0.13	paired passes, MC fails — random entries would do better
gap-and-go-stocks / donchian-stock-breakout	EQUITY	21/39	1.85	+1.67/+0.82%	0.79/0.66	1.00	MC passes, paired fails — gains track the window, not skill
adx-trend-scout	ETF	35	1.76	+0.10%	0.68	0.66	PF alone overrates it — why both tests are mandatory

Combined-book reproduction (Phase 1 of the adopted plan)

Correction after review. An earlier version of this page claimed "two real bugs" in the reference parser. That claim was wrong and is retracted. The reference reads Ken French's current *_CSV.zip files (comma-separated, first block = value-weighted monthly, data through 2026-08). My first run used the older TXT flavor (whitespace-delimited, cp1252, ends 2022-08) — the "bugs" were my data-source mistake, not defects in the reference code. Re-run with the correct CSV zips and the unmodified reference reproduces committed results.csv exactly:

Strategy	Ref Sharpe	Repro Sharpe	Verdict
Combined inv-vol book	1.2666	1.2666	exact · 8.85% / −20.45%
QualVal (FF hi-OP+hi-BM)	0.8589	0.8589	exact · data ends 2026-08
TSMOM 1/3/12 10%vol (pre-reg)	0.8563	0.8563	exact · 7.46% / −25.33%
TSMOM WF OOS 2016+	0.3148	0.3148	exact
PutWrite (CBOE PUT)	0.8413	0.8413	exact · 8.59% / −32.66%
Factor ETF blend 2013+	0.9557	0.9557	exact
SPY buy&hold	0.7770	0.7770	exact
Coin-flip control	−0.2829	−0.2829	rejected as control (PF < 1)

Max |Sharpe Δ| across all 8 rows: 0.0000. The reference numbers stand unchanged — nothing in grokbot needs a fix, and the earlier divergence was purely the data flavor (TXT ending 2022-08 vs CSV zips through 2026-08). The reference’s October 2008 inputs also check out against the CSV tables (OP Hi 30 −15.53%, BE-ME Hi 30 −21.13%, 50/50 blend after costs −18.35%).

Coordinated parallel work (same day, other agents/branches)

Money-ready plan Tasks 1–4, 9: per-card NOT MONEY-READY chips, ENTER/WAIT/AVOID tiers, real 3-scenario IV band replacing the fixed-multiple target, earnings-date graceful no-op, Wilson CI + n<20 hide, automatic kill banner (12 consecutive losses / WR<35% over 30) — TDD, 59 tests green at last report.
Part C: pick_scorer.py (Ranked Edge composite) + compounder_screener.py (QMJ/aristocrat screen, basket parameters).
Combined-book paper-trade branch (feat/combined-book-paper-trade, dabdbcbb69): research dataset harness, reproduction tests (43/43), frozen RULES.md with hash, spec "Draft for user review".

3. The plan (as adopted today)

Phase 1 — reproduce & calibrate. Done. Reference backtest reproduces within tolerance; coin-flip control correctly rejected; gates sort no-edge vs edge correctly (both Grok's harness and our own seeded tests).
Phase 2 — lock it in. Freeze pilot RULES + SHA-256 (done on the combined-book branch); every published card NOT MONEY-READY; no ENTER anywhere until a class clears the gate (renderer work shipped today).
Phase 3 — paper pilot, monthly. Record return, re-check hash, run gates, verdict: not-ready / ready / kill. Checkpoints: 12 months (live inside backtest range, else investigate), 24 months (kill rules start), 100 months (full go-live gate). First independent check-in: Nov 2.
Phase 4 — per-asset-class tests. One most-studied strategy per class, settings chosen on pre-2015 data only, tested 2016+. Pass bars: stocks (QMJ basket beats SPY Sharpe post-2016 + beats 95% of random baskets); bonds (trend+carry beats AGG/IEF); commodities (trend+carry Sharpe>0.3 post-2016 + crisis-positive); ETFs (fees still pass); volatility (defined-risk put-write, drop ≤ its 2008/2020 worst); crypto (delta-neutral funding carry, 3y+ after-cost). Failure protocol: one documented retry only (counted in trials); then drop, or hold as low-weight diversifier judged on drawdown reduction; regime-failed strategies (like TSMOM post-2015) stay as capped crisis hedges, never marketed as edges; if everything fails, publish honest passive 60/40/risk-parity.

4. Areas for future research

Wire Task 10 into the dashboard — the benchmark proof file exists; the system cards don't consume it yet (left untouched to avoid clashing with the renderer branch).
PEAD: shadow → production — the repo's only WF-verified strategy with zero live trades; flipping the wiring flag is the cheapest new-edge experiment available.
Funding-carry positive-regime gate — carry strategies were correctly killed in an inverted regime; a 7-day-average funding floor would let them re-enter legitimately.
TSMOM ETF sleeve → n≥100 — only live-positive sleeve (PF 1.76); needs scale to clear the CI-LB bar; watch the 2026 valuation-boundary overlay (halve at CAPE top decile).
Proof-layer upgrades — GRS test vs SPY (formal "adds value over buy-and-hold?"), split-conformal 90% return intervals on pick cards, O'Brien-Fleming sequential gate on the ETF pilot (anti-peeking), crowding proxy filter on the compounder basket.
Per-class Phase 4 sleeves — bond trend+carry, commodity roll yield, defined-risk vol-premium — each behind pre-registration + one-retry rule.
Ken French data hygiene — always use the *_CSV.zip flavor (comma-separated, current through the latest month). The TXT flavor ends 2022-08 and silently produces stale factor sleeves; it is not a valid substitute.

Verification ledger (this session)

pytest tests/test_benchmark_closed_ledger.py → 33/33, exit 0 (after fixing 3 test bugs found by the first red run — causes fixed, not asserted away; +5 tests added for the MC same-asset fix, incl. the coin-flip gate)
py_compile alpha_engine/benchmark_closed_ledger.py → OK
Task 10 smoke run on dashboard_data.json + live Yahoo prices → exit 0; report strict-JSON verified
Combined-book reproduction run.py → exit 0, unmodified reference code, Ken French *_CSV.zip data; 8/8 rows identical to committed results.csv (side-by-side table above)

Honesty notes. Nothing here is investment advice. No class is money-ready (0/9, the site's own verdict files say so). The combined book fails the 20% drawdown gate by 0.5pp full-period; it passes only 2016+. Backtests use proxy history (mutual-fund splices, index vs tradable instruments) and one stale factor file — the pilot exists precisely to replace backtest claims with live paper records. Changing any rule restarts the pilot clock.

Generated by the Freebuff agent session of Oct 19, 2026. Related: ← updates index · /audit dashboard.