Forward testing an EA: how long is long enough?

[FXS London Scalper]

After a backtest, a forward test on demo or small live is standard. The question is how long to run it.

Factors that decide:

• Trade frequency: an EA with 5 trades a week needs months; one with 20 a day needs less calendar time.
• Variety of conditions: a forward test in only quiet markets says little about news weeks.
• Expected edge: a small edge needs more trades to show itself.

A rough guide: at least a few hundred trades, or a few months including some major news weeks, whichever is longer.

What to compare:

• Average result per trade against the backtest for the same period.
• Distribution of results, not only the mean.
• Slippage and costs against assumptions.

If forward results are worse than backtest by more than costs explain, something's wrong in the backtest.

Write the stopping criteria down before starting, so the length of the test isn't decided by whether early results look good.

How long do you forward test?

[Amanhasnoname]

For a forward test I use a time rule and a trade rule together, whichever takes longer. My last EA was a mean-reversion system that traded about eight times a day, and I decided on 400 trades and at least one month with a full set of major releases. It took seven weeks, and I'm glad I didn't stop at the two good weeks at the start.

What I compare is the average result per trade in pips, the worst consecutive losing run, and the average slippage against what I assumed in the backtest. In my case the average result was about 30 percent below the backtest, almost all of it explained by slippage and a wider spread than I'd tested with. That was acceptable, so I went live.

I also wrote down the stop criteria on paper before starting: a drawdown of 12 percent on the demo balance, or an average result below zero after 200 trades. Having that in writing removed the temptation to tweak parameters halfway through, which I used to do all the time.