How the honest backtest works

There are two kinds of backtest on this site’s horizon, and they answer different questions.

	Browser Backtest	Zunder’s research backtests
Question	what would my rules have changed about these real trades?	would this strategy have made money?
Input	an address’s recorded fills, funding and orders	candles, a strategy, a fill model
Prices	the ones that happened	simulated fills
Runs	in your browser (crates/zunder-risk-wasm, being built)	on Zunder’s build box (zunder-sim, zunder-research)

The browser Backtest: replay, not simulation

It never invents a price. Every trade in both curves was a real fill; the guarded curve only takes less of it, or none.

The rules of the replay (crates/zunder-risk-wasm/src/replay.rs):

1. Actual equity = starting equity + closed PnL − fees + funding, as recorded.
2. Guarded entries are judged by the same function as Watch and the live market, against the guarded account. Refused: skipped. Resized: scaled.
3. Guarded exits follow the account’s exits in proportion. If the account closed half, the guarded position closes half.
4. Fees and funding of the guarded position are the recorded ones, scaled to its size.
5. A real RiskEngine observes guarded equity at every event. A daily loss stop blocks entries until the next UTC day; a drawdown halt blocks them for good. Guarded positions are closed at the account’s next fill price in that coin.
6. The audit judges every fill against the account as it really was, with a second engine on the actual equity.

Example

The account buys 2 ETH at 2,500 with a stop at 2,400 on 2,000 equity, later sells 1 ETH at 2,600 and 1 ETH at 2,450.

• Guard allows 0.38 ETH: 40 / (100 + 3) = 0.388…, with 12 bps of 2,500 = 3 as costs, rounded down to a lot size of 0.01 ETH (an assumption for the example).
• The guarded exits are half and half: 0.19 ETH at 2,600 and 0.19 ETH at 2,450.
• Guarded PnL: 0.19 × 100 − 0.19 × 50 = 9.50, before fees. Actual: 1 × 100 − 1 × 50 = 50.

The guarded curve made less on this trade. It also risked a fifth as much. The page shows both, and never hides the cases where Guard would have cost profit.

Honesty rules

These come from how Zunder treats its own research (CLAUDE.md, hard rule 8: “A backtest that looks great is a bug until proven otherwise”), applied to the browser tool:

• Costs included. Fees and funding as Hyperliquid recorded them.
• Assumptions on the page. Every result lists what the replay could not know (the list).
• No promise. “Would have lost less on this history” is a statement about the past, under stated assumptions. It is not a forecast.
• Any address. The page is prefilled with a public vault as an example, and invites you to replay any address, including ones where Guard would have cost profit.

Zunder’s research backtests

For its own strategies, Zunder applies stricter rules (docs/strategy-research.md): a strategy sees a candle only after it closed; a random-walk test and an independent second implementation of the backtest loop guard against look-ahead; parameters are tuned on one period, checked on a second, and tested once on a held-out third; every variant tried is counted, and the result is corrected for that count (deflated Sharpe ratio). Most ideas end as “rejected”. Those results are research notes, not product claims.