Parameter Sweep

Test a whole range of parameter values in one run — and read the results without fooling yourself.

Instead of asking "is RSI period 14 right?", a sweep runs the same strategy across a range of values — period 8, 10, 12, … 20 — in one dispatch, then ranks the results. It's the fastest way to see whether a strategy's edge is stable around its parameters or balanced on a knife's edge.

Running a sweep

  1. Open Sweep in the sidebar. You don't pick a saved strategy — choosing an axis picks the strategy family for you (SMA/EMA cross, RSI extremes, or MACD signal), and the sweep builds the base strategy from it.
  2. Choose the parameter axis to vary and its range. Studio sweeps one axis (up to 30 combinations); Quant can grid up to three axes at once (bounded at ~512 combinations).
  3. Pick market, timeframe, and period like any backtest, then dispatch. Progress streams in; results appear as each combination finishes.

Reading the results

The leaderboard ranks combinations, and the heatmap/sparkline grid shows how performance changes as a parameter moves. The shape matters more than the single best cell:

  1. A plateau — many neighboring values performing similarly — suggests a stable edge that survives small parameter changes.
  2. A lone spike — one great cell surrounded by losers — is the signature of overfitting. That exact value happened to fit the past.
  3. Check the out-of-sample column: a combination that only wins in-sample didn't generalize.

Overfit check and the luck baseline

Ranking N combinations and keeping the winner is multiple testing: even with zero real edge, the luckiest of N tries looks good. So every sweep result carries an Overfit check that deflates the winner's Sharpe by what the search alone would have produced. It runs on every sweep, on every plan — it is never gated.

  1. Luck baseline — the Sharpe the best of N tries reaches by chance alone. It is the Expected Maximum Sharpe (López de Prado, 2014): take the spread (standard deviation) of your runs' Sharpes and scale it by how far above the mean the best of N normally lands. More combinations, or more spread between them, raises the bar.
  2. Edge after selection — best Sharpe minus the luck baseline. Positive means something survived the search; zero or negative means the winner is explained by selection luck, not edge.
  3. When it can't be estimated it says so instead of disappearing. It needs at least 5 runs that actually produced a Sharpe, and some spread between them — runs with no trades can't be scored, so a wide sweep of dead combinations may still be unmeasurable.

Multi-symbol sweeps (Quant)

Quant can run one parameter grid across up to 5 symbols at once, ranked by a robust score — the mean return across symbols minus one standard deviation. A combination that wins on BTC but loses on ETH ranks below one that's merely decent everywhere: cross-asset consistency beats single-asset brilliance.

Finding past sweeps

The Sweep page opens on your history: every run you have made, newest first, each carrying the same overfit verdict its result page shows. Click one to reopen its leaderboard, or delete one you are done with — deleting a sweep leaves any strategy you saved from it untouched, and does not give the quota back. Every run also has a permanent URL (/strategy/sweep/<run id>) you can bookmark; it is private to your account, not a share link. The most recent 50 runs are kept per account.