Describe an edge in plain language. A team of AI agents codes it, verifies it, backtests it with real exchange fees, and hunts for overfitting — then paper-trades the survivors on live data. You write no code; no capital is ever at risk.
Every idea runs a rigorous gauntlet — real fees, out-of-sample holdout, overfitting checks. The platform is built to reject curve-fits, not celebrate them.
97% of ideas are abandoned automatically. That discipline is the product — you only ever look at what survived.
Every strategy comes with the analysis a quant desk would run — net-of-fee equity curve, risk metrics, trade-by-trade breakdown, and a return distribution.
Backtesting your own ideas takes months to do right. Copying a signal group means trusting a track record you can't see. Here's the difference.
| Stratmill | Do it yourself | Signal groups | |
|---|---|---|---|
| No coding required | |||
| Real fees, funding & slippage modeled | If you build it | ||
| Overfitting & holdout checks | Rarely | ||
| Full, verifiable track record | Opaque | ||
| Your ideas stay private | |||
| Time to first backtest | Minutes | Days–weeks | — |
Say your edge in plain language. A research agent turns it into a testable, fee-aware hypothesis.
A developer agent codes it; four verification layers catch look-ahead bias and broken logic.
Real maker/taker fees, funding, leverage and liquidation on historical data. Net-of-cost only.
Sensitivity, walk-forward, and a one-shot holdout. Overfits are abandoned automatically.
Survivors run live on the same engine as the backtest. No real capital, ever, in v1.
A strategy that only works at zero cost isn't a strategy. We charge the actual fee schedule of each venue — a round-trip taker costs ~0.10%, so the edge has to clear it.
| Venue | Instrument | Maker | Taker | Max leverage |
|---|---|---|---|---|
| Binance | USD-M Futures | 0.020% | 0.050% | 20× |
| Hyperliquid | Perps | 0.015% | 0.045% | 20× |
| Bybit | Linear | 0.020% | 0.055% | 20× |
| Binance | Spot | 0.100% | 0.100% | 1× |
| Deribit | Options | 0.030% | 0.030% | — |
Funding, slippage and market-impact are modeled too — PnL is always net.
Research, development, QA, analysis, risk and portfolio agents collaborate on every strategy — each an expert in its stage.
Four verification layers plus walk-forward and holdout validation. Deflated-Sharpe and cliff-parameter checks reject curve-fits.
Actual fees, funding, leverage, liquidation, slippage and market impact — modeled on real historical order-flow.
Subscribe to proven strategies — full track record visible, source code protected. Publish yours and earn 70%.
Paid workspaces learn from the platform's entire accumulated research — hundreds of tested strategies of hard-won signal.
Paper trading runs the exact same engine as the backtest, on live data — what you validated is what runs.
Every listing shows its real backtested track record; the source code stays with the publisher until you subscribe. You get a runnable copy in your own workspace. Share your own strategy and earn 70% of every subscription.
Browse the marketplaceUsage is metered in credits. Free forever to explore; paid plans unlock the platform's research corpus.
No. You describe an idea in plain language; the agents handle the quant work. You review results, not code.
No. Version 1 is research and paper-trading only — strategies run on live market data in a simulated account. There is no live execution with your capital.
Sensitivity analysis, walk-forward optimization with in-sample/out-of-sample tracking, a one-shot holdout, and deflated-Sharpe checks. Strategies that only look good in-sample are abandoned automatically.
Every backtest charges each venue's real maker/taker fees and models funding, leverage, liquidation, slippage and market impact. PnL is always net of costs.
A prepaid unit of platform usage that meters the AI compute your research consumes. Free plans include a monthly allowance; paid plans include more and never-expiring packs are available.
A good backtest is easy to fake. Every strategy is put through the same statistical tests a professional quant researcher would demand — designed to expose luck, not reward it.
Catches a Sharpe that only looks good because many variations were tried. Adjusts for the number of trials so lucky fits fail.
Catches parameters tuned to the past. Fit in-sample, scored out-of-sample across rolling windows — decay shows up immediately.
Catches overfitting to the whole dataset. A final untouched slice is tested exactly once — no second chances, no peeking.
Catches knife-edge parameters. Each input is swept ±20%; if a small change collapses the edge, it’s rejected.
Catches edges too thin to trade. Average per-trade return must clear real round-trip fees, or the strategy is abandoned.
Catches in-sample flattery. If out-of-sample performance falls too far below in-sample, the result is treated as noise.
Put your first idea through the pipeline in minutes — free, no card required.
Stratmill is a research and paper-trading tool, not financial advice or a broker. Backtest and paper results are hypothetical and not indicative of future performance. Figures shown are illustrative. Trading involves substantial risk of loss.