EA Optimization Best Practices: Finding Robust Parameters Without Overfitting

The Optimization Paradox

Optimization is a double-edged sword. Done right, it finds parameters that extract maximum edge from your strategy. Done wrong, it creates a perfectly curve-fitted EA that fails spectacularly in live trading.

The goal isn't finding the "best" parameters. It's finding parameters that are:

Profitable across different market conditions

Stable (nearby values also work)

Logical (make sense given your strategy)

Consistent in out-of-sample testing

Remember: If your EA needs exact parameters to work, it's not a strategy - it's a coincidence.

Genetic Algorithm Settings

MT5 Optimization Modes

Slow complete algorithm:

Tests every combination

Best for < 1,000 total combinations

Guarantees finding the global optimum

Can take days for complex EAs

Fast genetic algorithm:

Uses evolutionary selection

Finds near-optimal solutions quickly

Best for > 1,000 combinations

May miss isolated optimal points

Recommended GA Settings

`` Population size: 256 (balance speed vs coverage) Generations: Automatic (let it converge) Selection: Tournament Crossover: 0.9 (90% of new population) Mutation: 0.1 (10% random changes) ``

Optimization Criteria

| Criteria | Best For | |----------|----------| | Balance | Maximizing raw profit | | Profit Factor | Risk-adjusted returns | | Expected Payoff | Consistency per trade | | Drawdown | Minimizing risk | | Sharpe Ratio | Overall risk-adjusted performance | | Custom | Your specific goals |

Our recommendation: Optimize for Profit Factor or Custom criterion that balances profit and drawdown.

Parameter Selection

Which Parameters to Optimize

Good candidates:

Indicator periods (MA length, RSI period)

Entry/exit thresholds

Stop loss and take profit distances

Position sizing multipliers

Bad candidates (usually):

Magic numbers

Lot sizes (use risk-based sizing instead)

Broker-specific settings

Timing parameters (hour, minute)

Setting Parameter Ranges

Too narrow: Might miss better values Too wide: Wastes time testing impossible values

Example - RSI Period: `` Bad: 2 to 100, step 1 (99 values, many useless) Good: 10 to 30, step 2 (11 values, all reasonable) `

Example - Stop Loss (pips): ` Scalper: 5 to 20, step 5 Day trader: 20 to 60, step 10 Swing trader: 50 to 150, step 25 `

The Stability Rule

Test nearby values for each optimal parameter:

` Optimal MA period: 21 Test 19, 20, 21, 22, 23

If all are profitable: ROBUST If only 21 works: CURVE-FITTED ``

Walk-Forward Optimization

Walk-forward is the gold standard for finding robust parameters.

How It Works

1. Divide your data into segments (e.g., 6-month periods)

2. Optimize on segment 1

3. Test best parameters on segment 2 (out-of-sample)

4. Optimize on segments 1+2

5. Test on segment 3

6. Continue...

Example Timeline

`` 2019 Jan-Jun: Optimize (In-Sample) 2019 Jul-Sep: Test (Out-of-Sample) ✓ 2019 Jan-Sep: Optimize (In-Sample) 2019 Oct-Dec: Test (Out-of-Sample) ✓ 2019 Jan - 2020 Jun: Optimize (In-Sample) 2020 Jul-Sep: Test (Out-of-Sample) ✓ ... `

Walk-Forward Efficiency Ratio

` WF Efficiency = Out-of-Sample Profit / In-Sample Profit

> 0.5: Acceptable (strategy has real edge) > 0.7: Good (robust strategy) > 0.9: Excellent (very stable strategy) < 0.3: Poor (likely curve-fitted) ``

Walk-Forward Matrix (MT5)

MT5's Strategy Tester includes built-in walk-forward testing:

1. Open Strategy Tester

2. Select "Forward" period

3. Choose forward mode: "Half year", "Year", or "Custom"

4. Run optimization

5. Check "Forward" results in report

Robustness Testing

Monte Carlo Simulation

Test how your EA performs under randomized conditions:

1. Trade order randomization: Does profit depend on trade sequence?

2. Parameter perturbation: What if parameters are slightly off?

3. Data variation: How does noise affect results?

Use our Monte Carlo Calculator to stress-test your parameters.

Multi-Market Validation

A robust strategy should work on similar markets:

`` Original pair: EURUSD ✓ Test on GBPUSD: ✓ (should also work) Test on USDJPY: ✓ (should also work) Test on AUDNZD: ✓ (optional, different dynamics) ``

If it only works on EURUSD, you've likely curve-fitted to that pair's specific history.

Regime Testing

Test across different market regimes:

| Regime | Date Range | Your EA Result | |--------|------------|----------------| | Trending | 2017 | Should be profitable | | Ranging | 2018 | Should survive | | Crisis | 2020 Q1 | Should not blow up | | Recovery | 2020 Q2-Q4 | Should work | | High rates | 2022 | Should adapt |

Implementation Guide

Step-by-Step Optimization Process

Phase 1: Initial Exploration ``

1. Set wide parameter ranges

2. Run genetic optimization on 60% of data

3. Identify promising regions

4. Note parameter clusters that work

`

Phase 2: Refinement `

1. Narrow ranges around promising regions

2. Run detailed optimization

3. Select top 10 parameter sets

4. Test stability of each

`

Phase 3: Walk-Forward Validation `

1. Run walk-forward optimization

2. Calculate WF efficiency ratio

3. Reject if < 0.5

4. Keep sets with > 0.6 efficiency

`

Phase 4: Robustness Testing `

1. Monte Carlo simulation (1000 runs)

2. Multi-pair testing

3. Regime testing

4. Final parameter selection

`

Code for Custom Optimization Criterion

`mql5 double OnTester() { // Get standard metrics double profit = TesterStatistics(STAT_PROFIT); double drawdown = TesterStatistics(STAT_EQUITY_DD_RELATIVE); double trades = TesterStatistics(STAT_TRADES); double profitFactor = TesterStatistics(STAT_PROFIT_FACTOR);

// Reject if too few trades if(trades < 100) return 0;

// Reject if drawdown too high if(drawdown > 30) return 0;

// Reject if profit factor too low if(profitFactor < 1.2) return 0;

// Custom score: Profit * Profit Factor / Drawdown double score = profit * profitFactor / (drawdown + 1);

return score; } ``

Conclusion

Optimization is an art as much as a science. Here's your checklist:

Before Optimizing

Define what "success" looks like (criteria)

Set logical parameter ranges

Reserve 30% of data for out-of-sample testing

Know your strategy's theoretical basis

During Optimization

Use genetic algorithm for large parameter spaces

Optimize for risk-adjusted returns, not just profit

Check parameter stability (nearby values should work)

Run multiple optimizations to confirm results

After Optimizing

Walk-forward validate (WF efficiency > 0.5)

Monte Carlo test (95th percentile acceptable?)

Multi-pair test (works on similar markets?)

Regime test (survives different conditions?)

Red Flags

Only one parameter set works

Very specific values (SMA 47 but not 46 or 48)

WF efficiency below 0.3

Dramatically different results on similar pairs

Works only in specific time periods

The final test: If you wouldn't bet your own money on the parameters, don't deploy them.

Need help optimizing your EA? Get a free strategy audit or contact us for professional optimization services.