By Cree Lawson / Prismatic Algos
There is no universal “100 trades and go live” rule for an EA. The number depends on the net edge you need to detect and the variation between trade outcomes. A fixed count can serve as an order-handling pilot, but it does not prove a profitable strategy.
Define the net edge first
Start by defining the result of one completed trade after spread, commission and slippage. Average those net results to estimate expectancy. The familiar win rate is insufficient: 60% winners can still lose money if average losses are large. Before a demo run, lock the EA's code, parameters, symbols, risk and the minimum net expectancy that would justify deployment.
The NIST sample-size method gives a planning approximation for a two-sided mean test: N = (z₁₋α/₂ + z₁₋β)²(σ/δ)² . Here δ is the net edge to detect, σ is trade-result standard deviation, α is the false-positive rate and 1−β is test power. At 5% significance and 80% power, assuming independent, approximately normal outcomes and known variance, the illustrative requirements are:
| Net edge to detect | Approximate trades |
|---|---|
| 0.25 × trade-result standard deviation | 126 |
| 0.20 × trade-result standard deviation | 197 |
| 0.10 × trade-result standard deviation | 785 |
These are planning calculations, not thresholds endorsed for any specific EA.
Correlated trades and overlapping positions reduce the effective information in a raw count. Trying many EA variants and publishing only the best result inflates the chance of a false discovery. Unknown variance, heavy tails and market-regime shifts can increase the needed sample. The test must also span enough calendar time to encounter different conditions: 197 trades at one per week take about 3.8 years; at one per weekday they take about 39 weeks.
Observe execution prospectively
During the prospective demo test, log signals, orders, fills, realized costs, rejected orders and downtime as well as closed trades. A material parameter change starts a new forward-test window. At the end, report the start/end dates, number of independent trades, net expectancy with an uncertainty interval, drawdown, and how many candidate settings were examined. Check operational behavior and statistical evidence separately.
MetaQuotes' strategy-testing reference explains why modeling mode and tick quality matter in backtests. A prospective demo adds evidence about actual signals and broker-like execution, but it cannot reproduce every funded-account cost or liquidity effect. The decision to trade live also depends on risk limits and capacity, not only sample size.
Method note: The 126/197/785 figures are rounded-up substitutions into the NIST formula, not estimates from our strategy catalog. No fixed number of demo trades can establish an edge without the strategy-specific return distribution and a declared detection target.
