How can you create/find a truly profitable EA? - page 3
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
Hybrid or data-driven decision-making tools are definitely the safer path compared to blind automation.
Yes, algorithmic trading should never be done blindly. Always test on real tick data from your live OTC broker-dealer's or exchange's historic data with live spreads, latency/slippage, and/or swaps simulated/factored in─manually factored in if necessary. I prefer to run through at least 4000 round trades in any test. Again, this assumes that your code is fully Tester-compliant as well. I also like to artificially inflate all fees a bit in the test. I would rather be surprised by live overperformance than live underperformance based on the test report/statistics/probabilities. As with any software, the GIGO principle applies.
"[R]ead the room..."
I think there is an important distinction being missed in this discussion.
The real question may not be whether a human can “read the room” better than an EA. The more interesting question is whether that human judgement can be converted into a measurable market-state filter.
For example, instead of building an EA that is supposed to make money in every market condition, you can define several regimes — trending, ranging, high/low volatility, expanding/contracting volatility — and test separately whether the strategy actually has an edge in each one.
Then the EA does not need to “predict the market”. It only needs to answer a much simpler question:
“Is the current environment sufficiently similar to the environments in which this strategy has historically demonstrated an edge?”
That also changes how I would approach testing.
Real ticks are important if execution matters, but a beautiful tick-level backtest is still not evidence of robustness by itself. I would want to see something like:
in-sample vs out-of-sample performance
walk-forward testing
different market regimes
realistic spread/slippage/commission assumptions
parameter sensitivity rather than one optimal parameter set
Monte Carlo/randomized trade-order testing
and preferably a sufficiently large number of independent trades
The interesting part is what happens when the strategy is not trading.
If an EA makes money only during certain regimes and spends the rest of the time protecting capital, that may actually be a more realistic definition of a “profitable EA” than trying to force it to trade 24/5.
So I would agree with Victor on one point: a one-size-fits-all EA is probably unrealistic.
But I would disagree that this necessarily makes the solution discretionary trading. A human’s “read of the room” can often be decomposed into observable variables and tested.
The difficult part is proving that the decomposition works out-of-sample without simply creating another overfitted model.
That, to me, is the real challenge in building a profitable EA in 2026.
What would be more convincing to you: a single EA with a long live track record, or a portfolio of simpler EAs that each operate only in the regimes where they have demonstrated an edge?
I think there is an important distinction being missed in this discussion.
The real question may not be whether a human can “read the room” better than an EA. The more interesting question is whether that human judgement can be converted into a measurable market-state filter.
For example, instead of building an EA that is supposed to make money in every market condition, you can define several regimes — trending, ranging, high/low volatility, expanding/contracting volatility — and test separately whether the strategy actually has an edge in each one.
Then the EA does not need to “predict the market”. It only needs to answer a much simpler question:
“Is the current environment sufficiently similar to the environments in which this strategy has historically demonstrated an edge?”
That also changes how I would approach testing.
Real ticks are important if execution matters, but a beautiful tick-level backtest is still not evidence of robustness by itself. I would want to see something like:
in-sample vs out-of-sample performance
walk-forward testing
different market regimes
realistic spread/slippage/commission assumptions
parameter sensitivity rather than one optimal parameter set
Monte Carlo/randomized trade-order testing
and preferably a sufficiently large number of independent trades
The interesting part is what happens when the strategy is not trading.
If an EA makes money only during certain regimes and spends the rest of the time protecting capital, that may actually be a more realistic definition of a “profitable EA” than trying to force it to trade 24/5.
So I would agree with Victor on one point: a one-size-fits-all EA is probably unrealistic.
But I would disagree that this necessarily makes the solution discretionary trading. A human’s “read of the room” can often be decomposed into observable variables and tested.
The difficult part is proving that the decomposition works out-of-sample without simply creating another overfitted model.
That, to me, is the real challenge in building a profitable EA in 2026.
What would be more convincing to you: a single EA with a long live track record, or a portfolio of simpler EAs that each operate only in the regimes where they have demonstrated an edge?
As of reading the room ,I will give the example of 9/11 when a bot can't watch the news and hit the turn off button . Can you teach a bot to watch the news , imagine a martingale in that twin tower scenario 🤣.
Also an EA is a tool and you bring out different tools for a different job , the simpler EA's .
[A] one-size-fits-all EA is probably unrealistic.
Agreed. My best EA scalps based on a high resolution custom Renko chart. The EA is intentionally designed to target one instrument, and trades frequently enough to give the average swing trader a seizure. Not to mention, a manual trader would never be able to keep up with it even if she/he wanted to do so. Based on very specific statistics and probabilities collected from proper and accurate testing, the EA incorporates a time filter that operates on milliseconds and rather sophisticated position sizing. I would never suggest that any person try to do that manually, let alone allege that it's a one-size-fits-all EA.
Can you teach a bot to watch the news[?]
Certainly. Well really, you don't teach a bot─you code it. Many professional quants get paid big money to code bots that do exactly that, among other things. Speech-to-text and text-recognition have been around for decades. All a programmer has to do is write code that reads its output.
Are you aware that there are news wire services specifically for that purpose?
I enjoy discussing with a balanced guy that doesn't think he knows it all[.]
Another emotional outburst.
A forum mate recently asked me a similar question about my experience with EAs, and my answer was pretty simple: you won't find a ready-made profitable EA sitting on a forum or market, you have to build it around something you already know.
It completely depends on your strategy and the specific pair you actually understand. For example, if you've spent the last 3 years watching gold (XAUUSD) every single day, you know how it moves, how it reacts during sessions, and where it traps people. That familiarity is your actual edge.
When turning that into an EA, you just look at where you fail when trading manually:
Are you entering late because of emotion?
Are you moving your stop loss or cutting trades early?
Are you messing up your position sizing?
You build the code to solve those exact mistakes.
The rest is just keeping the logic simple. Don't add ten indicators or over-optimize 50 inputs in the tester to get a pretty curve. Focus on basic things that ruin systems in live conditions—spread spikes, slippage, and proper risk per trade. If your manual edge has clear rules and you know the pair well, turning it into a working EA is straightforward. If you don't understand the pair manually, code won't fix it.
Agreed. My best EA scalps based on a high resolution custom Renko chart. The EA is intentionally designed to target one instrument, and trades frequently enough to give the average swing trader a seizure. Not to mention, a manual trader would never be able to keep up with it even if she/he wanted to do so. Based on very specific statistics and probabilities collected from proper and accurate testing, the EA incorporates a time filter that operates on milliseconds and rather sophisticated position sizing. I would never suggest that any person try to do that manually, let alone allege that it's a one-size-fits-all EA.
Certainly. Well really, you don't teach a bot─you code it. Many professional quants get paid big money to code bots that do exactly that, among other things. Speech-to-text and text-recognition have been around for decades. All a programmer has to do is write code that reads its output.
Are you aware that there are news wire services specifically for that purpose?
Another emotional outburst.
A forum mate recently asked me a similar question about my experience with EAs, and my answer was pretty simple: you won't find a ready-made profitable EA sitting on a forum or market, you have to build it around something you already know.
It completely depends on your strategy and the specific pair you actually understand. For example, if you've spent the last 3 years watching gold (XAUUSD) every single day, you know how it moves, how it reacts during sessions, and where it traps people. That familiarity is your actual edge.
When turning that into an EA, you just look at where you fail when trading manually:
Are you entering late because of emotion?
Are you moving your stop loss or cutting trades early?
Are you messing up your position sizing?
You build the code to solve those exact mistakes.
The rest is just keeping the logic simple. Don't add ten indicators or over-optimize 50 inputs in the tester to get a pretty curve. Focus on basic things that ruin systems in live conditions—spread spikes, slippage, and proper risk per trade. If your manual edge has clear rules and you know the pair well, turning it into a working EA is straightforward. If you don't understand the pair manually, code won't fix it.