You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Hi Chris,
Good thread. Two things I haven't seen mentioned yet that helped me a lot.
1. Make sure the backtest itself is worth comparing against.
In MT4 a "Control points" or "Open prices" test can look completely different from live, so I only compare against Every tick tests and check the modelling quality in the report (90% is what MT4 gives with its own history). I also run the test with a spread close to my real average spread, not the fixed default, and then once more with a wider spread as a stress test. If the EA only survives with the tight spread, live will drift away from the backtest very early. In MT4 the tester also doesn't know your commission, so I add it to the spread or account for it manually – on a scalper that alone can turn a good backtest into a losing live account.
2. A switch-off line based on the EA's own statistics.
From the backtest I take the win rate and work out how long a losing streak is still normal. Roughly: with a win rate p and N trades, the longest expected losing streak is about log(N) / log(1/(1-p)). If live goes clearly beyond that – or beyond the longest losing streak in the windowed backtests – I stop and investigate, even if the drawdown is still below the backtest maximum. It reacts faster than drawdown, because drawdown depends on the lot size and on where in the cycle you started.
For the slow drift that nobody spots: I keep a rolling average of the slippage and of the time between the signal bar and the real fill over the last 20 trades. When that slowly gets worse, it's usually the VPS, the broker server or the spread at that hour – not the strategy.
Best regards.
Toyamato
Hi Chris,
Good thread. Two things I haven't seen mentioned yet that helped me a lot.
1. Make sure the backtest itself is worth comparing against.
In MT4 a "Control points" or "Open prices" test can look completely different from live, so I only compare against Every tick tests and check the modelling quality in the report (90% is what MT4 gives with its own history). I also run the test with a spread close to my real average spread, not the fixed default, and then once more with a wider spread as a stress test. If the EA only survives with the tight spread, live will drift away from the backtest very early. In MT4 the tester also doesn't know your commission, so I add it to the spread or account for it manually – on a scalper that alone can turn a good backtest into a losing live account.
2. A switch-off line based on the EA's own statistics.
From the backtest I take the win rate and work out how long a losing streak is still normal. Roughly: with a win rate p and N trades, the longest expected losing streak is about log(N) / log(1/(1-p)). If live goes clearly beyond that – or beyond the longest losing streak in the windowed backtests – I stop and investigate, even if the drawdown is still below the backtest maximum. It reacts faster than drawdown, because drawdown depends on the lot size and on where in the cycle you started.
For the slow drift that nobody spots: I keep a rolling average of the slippage and of the time between the signal bar and the real fill over the last 20 trades. When that slowly gets worse, it's usually the VPS, the broker server or the spread at that hour – not the strategy.
Best regards.
Toyamato
Hi Chris,
I think comparing trades side by side is a better way to spot issues than looking at PnL alone. A few bad weeks wouldn't be too concerning, but missing trades or completely different entry times would definitely raise some questions.
I'd check the execution logs, spreads, and latency first to see where the differences are coming from. Running the EA on a stable trading fx VPS might help reduce connection-related issues too.
As for when to stop the EA, setting a maximum drawdown limit before going live seems like a sensible approach. That way, you're following a clear rule instead of reacting to every losing streak.
Example from an EA I looked at, not one of mine. Same EA, backtest vs the seller's own live signal.
I think comparing trades side by side is a better way to spot issues...
Well said. If you don't want to have to sort through the log manually for entry times, see:
Hi Chris,
I do almost the same thing. I run the tester from the first live trade, on the same broker and symbol, and match the trades by entry time. On my last check 12 out of 12 trades lined up, with only small price differences. I also reset the count after every update, I agree that a new version is a new EA.
About the line, I split it into two kinds of problems because they need different answers:
1. Execution problems: the trades themselves are different
A trade on one side only, a different lot, or an SL/TP in a different place. For me even one unexplained case is enough to stop and look. Usually it's something technical: filling mode, a stops level rejection, the market being closed, or contract size on a different symbol suffix. This isn't noise, so I don't wait for more data.
2. Performance problems: same trades, the market just isn't paying
Here I don't compare against the backtest of the same weeks, but against the worst the backtest has ever done. I write down the historical max drawdown and the longest stretch without a new equity high. Live is still "normal" while it stays inside those two numbers. I would switch off if live drawdown goes clearly past the historical max (for me around 1.2–1.5x), or the flat period lasts much longer than anything in the history.
A few weeks below the backtest says very little for a swing system. Some months there are only a handful of trades, so one missed TP changes the picture. That's why I would rather look at whether the same trades happen than whether the profit matches.
One extra thing: I check swap and commission separately. On gold, holding long for several days can change the result a lot even when the entries are identical, and the tester doesn't always use the same swap the broker charges today.
Best regards,
Henry
On Christian’s point that one backtest gives you only one drawdown: cutting the history into windows with the same number of trades as the live sample gives you a distribution instead of a single figure.
For each window I record two things, the longest losing streak and the deepest drawdown in R, and fix the switch-off line at the worst 5% of each before going live.
Pairing them matters, because nine losses at 0.5R and nine at full size aren’t the same signal, and an EA that starts losing bigger rather than more often won’t trip a streak counter.Reshuffling the trade order, Monte Carlo style, widens the range further, although it assumes trades are independent and clustered losses aren’t. Execution differences, as Henry says, shouldn’t wait for any threshold.
Eduardo
On Christian’s point that one backtest gives you only one drawdown: cutting the history into windows with the same number of trades as the live sample gives you a distribution instead of a single figure.
For each window I record two things, the longest losing streak and the deepest drawdown in R, and fix the switch-off line at the worst 5% of each before going live.
Pairing them matters, because nine losses at 0.5R and nine at full size aren’t the same signal, and an EA that starts losing bigger rather than more often won’t trip a streak counter.Reshuffling the trade order, Monte Carlo style, widens the range further, although it assumes trades are independent and clustered losses aren’t. Execution differences, as Henry says, shouldn’t wait for any threshold.
Eduardo