Dashboard → https://amindraa05.github.io/trend-gated-megacap/
A point-in-time backtest of a concentrated US large-cap equity rule over 1996–2026, published together with the audit trail behind it. This repository holds the evidence. The engine and the live execution code are kept in a private repository.
Hold the fifteen most-traded S&P 500 constituents as they stood at the time. Cap any single name at a quarter of the book. Replace names monthly, but never trim a winner back to its target weight. Scale total exposure to a 30% volatility target. Hold cash whenever the market closes below its 200-day average. No leverage, no shorts.
| Window | CAGR | Max drawdown | $10,000 becomes |
|---|---|---|---|
| 2000–2026 | 19.71% | −43.5% | $1,186,129 |
| 2000–2009 | 17.66% | −35.4% | — |
| 2010–2017 | 24.33% | −16.5% | — |
| 2018–2026 | 17.51% | −43.5% | — |
| QQQ 2000–2026 | 8.57% | −83.0% | $88,682 |
| QQQ 2018–2026 | 19.94% | −35.1% | — |
The bold rows are the point. Since 2018 the index has beaten this rule on both return and drawdown. The advantage came from 2000–2009, when QQQ returned −6.76% a year and this rule returned +17.66%. Whether that matters depends entirely on whether another lost decade arrives.
Every claim below maps to a file in evidence/. Full inventory in
AUDIT_EVIDENCE.md.
- Accounting identity recomputed from raw vendor files — worst error 0.046 bps
- Point-in-time membership checked against 18 real index additions and removals — 0 early, 0 late
- Delisted companies included from the start; residual bias bounded at 9.3% of position slots
- Deflated Sharpe across 1,918 trials — 0.9995, p = 0.00055
- 1,000 random-pick placebo portfolios — the real result sits at the 100th percentile
- Configuration selected out-of-sample by anchored walk-forward, landing on the family median rather than the in-sample champion
- Every configuration ever run, winners and failures —
evidence/ledger.csv, 1,981 rows
Seven defects were found by these checks — two in the engine, five in the test harness itself. All are documented rather than quietly corrected.
- It has never been traded. No live or paper fill evidence exists. Every fill in every figure is an assumption about market-on-open execution.
- The design space was explored on the same history it is measured against. Walk-forward controls which configuration was chosen, not which ideas entered the search.
- One realised path of market history, resampled.
- Fifteen names. The worst single-name overnight move cost 19.5% of NAV in a day.
docs/ the dashboard served by GitHub Pages
evidence/ every result table and raw log the audit claims point to
evidence/ledger.csv all 1,981 configurations tested
AUDIT_EVIDENCE.md claim-by-claim inventory, including the gaps
Source data (EODHD end-of-day and point-in-time index membership) is licensed and is not redistributed.
Not investment advice. Not a product. A research record, published with its failures intact.