How we backtest every strategy

The engine

Backtests run through Keel’s own execution engine — the same code path that places live orders. A strategy you backtest and a strategy you deploy run through the identical logic, so what you see historically is what the engine would actually have done, not an idealised approximation. Fills, rebalancing, and funding accrual are modelled bar by bar.

The data

Real Hyperliquid perpetual market data over 2024-08-15 → 2026-07-05 — roughly two years spanning both strong uptrends and deep drawdowns. Prices, funding rates, and the tradable universe are the actual historical values, not synthetic or survivorship-filtered series.

Costs

Every headline number is net of costs. We charge each trade 4.5 bps of fees + 4.5 bps of slippage (a 9 bp round trip), and funding is included in every combined figure — a strategy that shorts a high-funding perp collects it; one that pays funding is charged it. There is no version of these numbers that ignores trading friction.

What we deliberately don’t do

• No cherry-picking. The headline on each page is the strategy’s default settings — never the single best-looking cell in the grid. When one setting happens to score higher, we say so and still lead with the robust default.
• Every setting shown, including the losers. The config explorer publishes the full grid — the settings that lost money are right there next to the ones that worked. Finding the bad configs yourself is the point.
• Drawdowns get equal billing. The worst historical dip is shown next to the return, in percent and in dollars at a reference balance. No return is presented without its risk.
• No “earn X%” claims. These are historical simulations on real data, not predictions or investment advice. Past results don’t guarantee future performance.

Keeping it current

The library is re-backtested on a monthly cadence and re-imported, so the numbers stay current. Each page shows its data as-of date; if a refresh ever lapses, the page flags itself as stale rather than quietly showing old numbers.