How to Optimize an Expert Advisor in MT5 Without Overfitting It to the Past

What optimisation actually does

The Strategy Tester runs your EA many times over the same history, each time with a different combination of the inputs you have ticked for optimisation, and ranks the runs by a result you choose. That ranking is a search, and a search will always return a winner. The winner tells you which combination fitted this particular stretch of history best. It does not tell you whether the combination reflects something the market will keep doing. The gap between those two statements is the whole subject of this guide, and it has a name: overfitting.

To see the problem, imagine optimising a rule that has no meaning on prices that are pure noise. Try enough combinations and one of them will draw a smooth, rising curve. The more settings you try, the better the best one looks, and the less it means. Move the sliders in the calculator below to see how large the best-looking result from pure noise gets as the search widens.

The MT5 tester's optimisation options

In the tester you choose an optimisation mode, a criterion and an optional forward period. MetaQuotes documents all of them in the help pages on strategy optimisation. The table summarises the choices that matter most.

Optimise fewer inputs, on purpose

The number of combinations grows by multiplication. Five inputs with ten values each is 100,000 combinations; ten inputs is ten billion. The genetic mode hides that cost by only sampling, but the risk of luck grows the same way. The practical rule is to optimise the two or three inputs that define the strategy's logic, such as a lookback period or a threshold, and to leave the rest at sensible fixed values.

• Fix what is not strategy. Lot size, magic number, comments and slippage tolerance are not edges. Do not optimise them.
• Use coarse steps. Testing a 20-bar lookback against 21 and 22 teaches nothing. Step in ranges that matter, such as 10, 20, 30, 40, 50.
• Set ranges from logic, not from the data. A moving-average period of 3 or 300 is outside what the idea was meant to use. Keep the range sensible before you look at results.
• Respect the data you have. An EA that makes 30 trades in five years does not have enough trades for an optimisation to mean anything.

Pick a plateau, never a spike

After an optimisation, look at the results as a landscape, not a leaderboard. In the optimisation tab you can switch to a graph of two inputs against the result. What you want to see is a broad region where the result is good, so that nudging an input a little in any direction still works. That is a plateau. What you do not want is a single bright point surrounded by weak results. That is a spike, and it usually means the optimiser found a coincidence.

A plateau does not prove an edge, but a spike nearly proves the lack of one. The plateau also gives you something practical: because the neighbours work, small changes in market behaviour, spreads or execution will not break the strategy. For the one-dimensional version of the same idea, see how a robust setting looks in a walk-forward test.

Choose the criterion that rewards the right thing

Optimising for 'Balance max' means finding the combination that ended with the most money. That is the criterion most likely to reward one lucky, huge trade and the deepest drawdown, because large risk can produce large results on a short history. Better choices reward profit relative to risk. Recovery factor divides net profit by maximum drawdown, so a setting that makes money with a modest dip beats one that makes a little more and nearly blew up. A custom criterion lets you do more, for example penalising results with fewer than a minimum number of trades, or with a drawdown above a limit.

A repeatable process

1. 1

   Step 1: 1. Write the hypothesis first

   State in a sentence why the strategy should work: 'breakouts from the Asian range carry on during London'. If you cannot, you are searching, not testing.

2. 2

   Step 2: 2. Split the history before you look

   Reserve the most recent quarter to a third of the data as out-of-sample, and do not touch it while optimising. The Forward setting does this for you.

3. 3

   Step 3: 3. Optimise two or three inputs with coarse steps

   Use the genetic mode first to find the area, then the complete mode on a narrow range if you want detail.

4. 4

   Step 4: 4. Read the landscape

   Look at the two-input graph. Reject spikes. Note the extent of any plateau.

5. 5

   Step 5: 5. Pick the centre of a plateau

   Choose a value inside the broad region, not its edge or its single best cell.

6. 6

   Step 6: 6. Test it out of sample

   Run the chosen settings on the held-back data. Expect it to be worse than in-sample; the question is how much worse.

7. 7

   Step 7: 7. Stress it

   Add spread and slippage, shift each input by 10 to 20 percent and run on a different symbol or period. A robust system degrades gently.

8. 8

   Step 8: 8. Forward test on a demo

   Weeks of live data on your own broker, same version. See forward testing on a demo.

The mistakes that fool people most

• Peeking at the held-back data. Every time you adjust after seeing the out-of-sample result, it stops being out of sample.
• Optimising on one regime. A test over one long uptrend makes every trend system look good. Include a trend, a range and a shock.
• Trusting the best of many runs. The more variants you test, the more the winner is selection, not skill.
• Ignoring costs. Optimised scalping settings can look excellent at zero spread and fail at the broker's real one.
• Re-optimising constantly. Changing settings every few weeks to match recent results is chasing the noise. If you re-optimise, do it on a schedule and test the change out of sample.
• Believing a vendor's optimised set blindly. A set file is someone else's best fit. Check it on your broker's data and your own date range before you trust it. Our set file guide explains what is inside one.

The same discipline applies to buying. When we review an EA in our EA verdicts, a missing out-of-sample test or a result that rests on one window lowers the evidence score, because it is the first thing that fails when the market changes. For the theory behind how a backtest can mislead, the paper cited below is a readable account.