How to do if the EA always fails? But backtesting are fine

New comment
 
The buyers are the loser and helpless from MQL stuffs
 

If back testing works (as in it makes a profit) and live trades don't make a profit, then you're overfitting and/or doing something wrong with the test. Presumably you're testing within MetaTrader (as you give basically zero details for us to assist), in which case the most obvious mistake is not understanding how MetaTrader back test settings work, namely that "tick data" is made up data from OHLC values, and "tick data based on real ticks" actually downloads the data and uses it. You also need to make sure the data is actually good, if your modeling quality is like 30% then the back test is 100% useless. You'll want that at 99.9% or more.

If you're going to use EA's it's not as simple as add the EA to a chart and watch the money roll in. Magic Money Machines don't exist. At best an EA might trade a certain time frame for a few pairs, it's not going to just print money because you decided to buy a $30 piece of software and toss $100 into an account.

You got previous posts about "many sellers are willing to cheat buyers by manipulating backtesting and we as buyers lose a lot of money when buying EA. what can we do to prevent fraudsters?" - Half the problem is most people selling EA's don't actually know what they're doing. The other half of the problem is I'm pretty sure your expectations don't line up with reality. In your world if the EA doesn't print money on every time frame and every pair, it's fraudulent. An EA is just automating a task, and it's far easier to automate profitable trading on only EURUSD H1 than it is to somehow make it work on every pair, during trends, during reversals, etc. etc. Most human traders can't trade every pair and every time frame profitably, so expecting that from automation is a bit absurd. Find the guy who can trade one pair and one time frame really well, automate his/her strategy, and now you'll have a profitable EA.

 
James McKnight #:

First of all, I want to thank you for taking the time to write such a well-thought-out response. It’s been a while since I’ve come across something with so many solid facts presented so clearly. I agree with much of what you’ve said about overfitting and the importance of data quality in backtesting. That said, I believe there are a few key points worth further clarification:

  • Data Quality: You’re absolutely right that "tick data" can be synthetic when using OHLC mode, but even with "real tick data," its quality can vary depending on the data source and how missing ticks are interpolated. Obtaining accurate and complete data is essential, and achieving that 99.9% quality isn’t always as straightforward as it seems.
  • Backtesting vs. Live Trading: Beyond overfitting, factors like slippage, latency, and order execution conditions can significantly affect live profitability. These variables are often overlooked during backtesting but play a crucial role in real-world performance. It’s important that EAs account for these aspects to give a more accurate representation of their potential in live conditions.
  • Expectations from EAs: I completely agree that there’s no such thing as a “magic money machine.” However, I also believe that an EA that only works profitably on a single pair and timeframe is likely too limited. While a simple, focused approach can be effective, if an EA only performs well under very specific conditions, it may be overfitted or lack the robustness to handle different market scenarios.
  • Preventing Fraud: I share your view that a significant issue is that many sellers don’t fully understand what they’re offering. However, there is also a portion of the industry that exploits buyers’ lack of knowledge by using tactics like excessive optimization or manipulating backtests to create a false impression of profitability. Buyers should always demand forward testing results, Monte Carlo analysis, and a performance review across diverse market conditions to validate the effectiveness of an EA in real-world situations.

In conclusion, it’s not about expecting an EA to work across all pairs and timeframes, but about requiring transparency, rigorous testing, and a realistic approach when evaluating these automated systems.

 

If the EA is tick sensitive, then you must do due diligence and backtest on the real ticks model. Regardless of what the author said about it, you should obviously test it to prove that it's not tick sensitive if the author alleged that the system isn't tick sensitive. Some EAs will only work when the market is trending which means that not all EAs are to be left constantly on a chart. Only some EAs have more intelligence and can handle all market conditions (and can also apply risk management).

Some unpopular information that you might need to know - volatile bid and ask prices will hit your predictable narrow stop loss 99% of the time, and a tight trailing stop will especially destroy your account. You will see that a tight trailing stop works wonderfully on fake OHLC ticks. You really need to be careful about a trailing stop, and it might not be obvious at first.

 
James McKnight #:

If back testing works (as in it makes a profit) and live trades don't make a profit, then you're overfitting and/or doing something wrong with the test. Presumably you're testing within MetaTrader (as you give basically zero details for us to assist), in which case the most obvious mistake is not understanding how MetaTrader back test settings work, namely that "tick data" is made up data from OHLC values, and "tick data based on real ticks" actually downloads the data and uses it. You also need to make sure the data is actually good, if your modeling quality is like 30% then the back test is 100% useless. You'll want that at 99.9% or more.

If you're going to use EA's it's not as simple as add the EA to a chart and watch the money roll in. Magic Money Machines don't exist. At best an EA might trade a certain time frame for a few pairs, it's not going to just print money because you decided to buy a $30 piece of software and toss $100 into an account.

You got previous posts about "many sellers are willing to cheat buyers by manipulating backtesting and we as buyers lose a lot of money when buying EA. what can we do to prevent fraudsters?" - Half the problem is most people selling EA's don't actually know what they're doing. The other half of the problem is I'm pretty sure your expectations don't line up with reality. In your world if the EA doesn't print money on every time frame and every pair, it's fraudulent. An EA is just automating a task, and it's far easier to automate profitable trading on only EURUSD H1 than it is to somehow make it work on every pair, during trends, during reversals, etc. etc. Most human traders can't trade every pair and every time frame profitably, so expecting that from automation is a bit absurd. Find the guy who can trade one pair and one time frame really well, automate his/her strategy, and now you'll have a profitable EA.

I agree with this . 

Many a times the seller thinks they found something and are oblivious to overfitting concepts, and it does not matter if the strategy was built with a neural net , chat gpt , pen and paper testing , the strategy tester or they dreamt about it. Everything can overfit

The problem the buyer or anyone can't see is the duration of the training set.

If i deploy a strategy on the entire history of an asset and then claim it was trained up to 2023 and validated and tested in 2024 and 2025 there is no way on earth for anyone other than the seller to validate this claim.

So the buyer will go ahead and do the little forward testing tests and believe the ea is the holy grail.

This applies to manual human built no ai strategies as well. When a product reads "i've been developing this strategy for 5 years" you should interpret that as "i have possibly been overfitting this for 5 years" 

 
Lorentzos Roussos #:

I agree with this . 

Many a times the seller thinks they found something and are oblivious to overfitting concepts, and it does not matter if the strategy was built with a neural net , chat gpt , pen and paper testing , the strategy tester or they dreamt about it. Everything can overfit

Even sometimes an EA that shows an impressive backtest on the real ticks model could make almost no money at all on the live market, and not only that...but it could fail. So the only thing you can really trust is a signal that reveals positive live trading results (over one year minimum).

 
Conor Mcnamara #:

Even sometimes an EA that shows an impressive backtest on the real ticks model could make almost no money at all on the live market, and not only that...but it could fail. So the only thing you can really trust is a signal that reveals positive live trading results (over one year minimum).

exactly yes .

 
Kodj007:
The buyers are the loser and helpless from MQL stuffs

Only purchase EAs that have a signal, only run the EA on the same broker account type as the signal .. lots of EAs in the top 50 without signals .. have you ever asked yourself why ? its your money, DYOR.

.

Now, for the fun bit .. there is no such thing as a centralised FX data feed .. data is made up from liquidity providers (LPs), a broker can have 1 or more LPs, they can aggregagte feeds also .. so, even if you have the same account type on the broker, your feed could be slightly different and get differing trades .. also .. someone could eat up the liquidity meaning your trades are filled further away from your expected entry point and get a different entry ( slippage )

.

Also, have a google for A-Book and B-Book FX Brokers,  your broker could be both (Hybrid) and depending on how you chose to trade they may select whether you are A-book or B-book and, they may even take the opposite side of your trade so when you lose ( and most retail traders do ) they win .. a bit of light reading :)

.

Everybody thinks that if they back test using high quality tick data from dukascopy and such that it will translate into great live trades .. nope .. backtesting may give an idea that the strategy might be profitable .. but no crystal balls .. they only way you really know is to trade it .. hence the signals accounts .. 

.

my 2c, been doing this long enough to see all sorts .. 


all the best  MT

 

My Experience with Signals and Expert Advisors (EAs)

In today's environment, having a signal alone is no longer enough to "prove" the quality of an EA. While it can provide reassurance, it doesn't guarantee long-term success.

The real question is: what are you looking for? In my opinion, even a great EA only does 50% of the work. The other 50% depends entirely on the user: expectations, risk management, and discipline. Without proper practices, failure is very likely.

About risky strategies
Most EAs (especially in the top rankings) use aggressive methods to boost performance quickly:

  • Grid systems

  • Martingale

  • Extremely low risk-to-reward ratios
    Sometimes all three at once.

This doesn’t mean they can’t be used — but it depends on how they are used.

A few good practices to consider:

  • Once the account doubles, withdraw your initial capital

  • Continue trading only with the profits

  • Make regular withdrawals (weekly or monthly) to secure gains
    This type of discipline is key when working with high-risk strategies.

On the other hand, an EA that doesn’t use grid or martingale strategies will likely grow slower, but be more stable and can be left running without constant withdrawals.

Lastly, beware of backtests showing a profit factor above 10 — this is often a sign of over-optimization, and not a realistic performance indicator for live trading.

New comment