trading

A semi-automated trading system implementing Stan Weinstein's stage-analysis methodology. See docs/design/ for the system design and the Weinstein book reference.

The results below are regenerated mechanically by trading/backtest/readme_toplines. Two marker-delimited blocks are maintained, each overwritten on regeneration — do not hand-edit between the markers:

• the deep-headline block — the heavy multi-decade broad-universe results-of-record, rendered from the pinned records in dev/backtest/deep_headline_records.sexp (these runs use an out-of-repo warehouse and are not recomputed by CI);
• the light-reference block — recomputed by running the reference strategies end-to-end (period derived from actual bar coverage in the CSV store).

Deep multi-decade headline (results-of-record)

Result	Total return	Max DD	Trades	Win rate	Period
Weinstein top-3000 (promoted config)	+8,689%	30.3%	1,170	38.4%	2000-01-01 -> 2026-06-26
SPY total return (same window, comparator)	+706%	—	—	—	2000-01-01 -> 2026-06-26
Pre-bundle record (superseded)	+7,914%	32.3%	1,187	—	2000-01-01 -> 2026-06-26

Provenance (scenario / commit / date):

• Weinstein top-3000 (promoted config) — test_data/backtest_scenarios/staging-leverf-28y/top3000-2000-2026-rcb-f000.sexp @ 6a2d9b426 (PR #2047 — promoted bundle: w30 + virgin-crossing + floors-zero) (2026-07-23)
• SPY total return (same window, comparator) — n/a — dividend-adjusted SPY buy & hold @ DEEP_RESULTS.md record-of-record standing comparator (2026-07-14)
• Pre-bundle record (superseded) — test_data/backtest_scenarios/staging-record-convention/top3000-2000-2026-record-convention.sexp @ 0a2e4562d (PR #1960, Run D, dedup-v2 warehouse; DEEP_RESULTS record-of-record 2026-07-14) (2026-07-14)

Basis: mark-to-market, including open-position marks on a few concentrated fat-tail winners and (unless a liquidity overlay is armed) untradeable illiquid names — NOT bankable as realized P&L. The honest read is vs the index over the same window and realized-vs-MTM. Full pins + caveats: dev/backtest/DEEP_RESULTS.md.

Top-line results

Pinned testing period: 1998-12-22 -> 2026-06-12 (common bar coverage of SPY, BRK-B, and the nine original SPDR sector ETFs).

Strategy	Total return	CAGR (%/yr)	Notes
SPY buy-and-hold	+888.9%	+8.7%	buy & hold, dividend-adjusted close
BRK-B buy-and-hold	+1132.4%	+9.6%	buy & hold, dividend-adjusted close
SPY-only Weinstein	+408.0%	+6.1%	Spy_only_weinstein, 30-week investor MA, long/flat
Sector-ETF Weinstein	+528.9%	+6.9%	Sector_rotation_weinstein k=3, 30-week investor MA, RS vs SPY

Regenerate: dune exec trading/backtest/readme_toplines/bin/readme_toplines.exe -- --readme README.md (run inside the dev container).

Deep multi-decade backtests

The table above is the light reference set (auto-regenerated). For the heavy multi-decade broad-universe runs (top-3000 PIT-1998 and sp500-515, 1998/2000–2026, long-only + long-short, with the realized-vs-MTM and liquidity caveats), see dev/backtest/DEEP_RESULTS.md — the results-of-record, each row pinned to its scenario sexp and measurement commit. Those runs are experiment-only (reproducible locally against the deep data dir; skipped on GHA).

Margin & trade-safety model

Short selling and tradeability are modeled against real broker/regulatory rules (Reg-T 150% short collateral, FINRA 4210 maintenance, daily borrow fee, a $17 short-price floor for the 30% maintenance tier, force-liquidation/halt, and a default-off liquidity-realism overlay with an entry $-ADV gate + held-degradation exit). All controls are default-off no-ops — defaults reproduce a frictionless long-only engine bit-for-bit. See docs/design/margin-safety.md for the full mapping to broker requirements.