THE FINDINGS SYSTEM
Every test ends in one of three findings. Reading them is RAMP 02.
// THE THREE STATES
The strategy beat its benchmark on BOTH compound growth (CAGR) and risk-adjusted return (Sharpe) over the whole window, and in the walk-forward window after the split year it did not fall behind the benchmark on both. Both axes over the window, and not behind on both after the split, or it is not DATA VALIDATED.
It beat the benchmark on one axis only: more return with worse risk, or better risk with less return. Or it beat the benchmark on both over the whole window but fell behind it on both after the walk-forward split. You decide whether the trade is worth it.
The rule did no better than the benchmark. The result maps where the edge is not, and your AI reports it in those words.
// THE WALK-FORWARD SPLIT
Every stock backtest is split chronologically at the OOS SPLIT YEAR (2015 unless you pick another). The same criteria run unchanged through the whole window, and the result shows the years before the split and the years from it beside the whole run; nothing is re-fitted at the split. Random splits are not offered, by design: shuffling time leaks the future into the past, and a leaked backtest is a fiction with good numbers.
// THE METHODOLOGY, STATED IN FULL
- Delisted companies. Companies that later delisted are only partly in the market-wide backtest universe from 1998: on 2026-11-02, 12,521 of the 12,841 delisted operating companies with prices from 1998 (98%) could enter it, and each market-wide result states the share it measured. Earlier years hold survivors only.
- Fundamentals by filing date. Fundamentals are joined by filing date. About 9%
of quarters (44,745 of 490,372 on 2026-10-20) carry an estimated filing date, the
company's usual reporting delay or 45 days (90 for a fiscal Q4), marked
filing_date_imputed; none is dated on the quarter-end. A quarter keeps its first-reported figures once it is 30 days past its filing date; later restatements are stored beside it, not over it. - No look-ahead. Positions lag signals; formation-only filters; the forward return must span exactly one calendar period.
- Gross returns, costs beside them. Results are gross of costs by default, and
every result says so. Every
run_backtestresult also carries a cost model on its measured monthly turnover: net growth, Sharpe and worst fall at 5, 10, 25 and 50 bps per side. Taxes are not modelled, and real results would be lower than the gross figures. - Complete record. Every period, in the result. A backtest is not a promise of returns.
That is the whole system. When your AI presents QuantGPT numbers, it repeats this methodology, because the manifest it receives on connect tells it to.