In November 2024 I had the best backtest I'd ever seen. My scalping EA, run on GBPJPY M5 from 2020 through 2024 in the MT5 strategy tester, showed a 340% return with a max drawdown of 12%. I printed the equity curve and stuck it on my wall. Then I put $3,000 live behind it. Six weeks later the account was at $2,110 — down $890 — trading the exact same code. Same EA, same inputs, same broker. The backtest wasn't lying, exactly. It was answering a different question than the one I asked. That expensive lesson is why this mt5 strategy tester guide accurate backtests article exists: to make sure the test you're running measures reality, not fantasy.
These days I validate every backtest result against the DarkWolf MT5 Signal Calculator before I trust a single number the tester prints. Here's the full workflow.
Running My EA Through the Calculator Before Backtesting It
Before I even open the strategy tester, I run the EA's core trade through the calculator. Why? Because the tester will happily report a 340% return built on position sizes that would have margin-called a real account in month two. The calculator is my reality anchor.
| Check | Calculator input | Output | What it told me |
|---|---|---|---|
| GBPJPY scalp | $3,000 · 2% risk · 15-pip SL | 0.55 lots, $60 risk | EA's fixed 1.0 lot was 3.6% risk, not 2% |
| News spike trade | $3,000 · 2% risk · 15-pip SL + 8 slip | 0.36 lots, $60 risk | Real conditions cut size by a third |
| Overnight hold | $3,000 · 1% risk · 40-pip SL | 0.10 lots, $30 risk | Swap costs needed separate accounting |
The first row exposed the real bug in my "amazing" backtest: the EA traded a fixed 1.0 lot, which the calculator showed was 3.6% risk per trade on my $3,000 account — not the 2% I thought. The tester doesn't warn you about that. It just compounds the fantasy.
What the Tester Actually Simulates (and What It Fakes)
The strategy tester has four modeling modes, and picking the wrong one is where most fantasy backtests begin. "Every tick" generates synthetic ticks inside each M1 candle — plausible but fake. "Every tick based on real ticks" uses your broker's actual tick history — this is the only mode I trust for scalpers. "M1 OHLC" and "Open prices only" are faster but useless for anything that trades inside the candle.
Here's what the tester fakes even in the best mode. Limit orders fill at the touched price — but in real markets, when price wicks through your limit by a single point during news, you often don't get filled at all; the liquidity isn't really there. The tester also assumes your 5-lot market order doesn't move the price. On EURUSD, fine. On an exotic pair at 3 AM, your own order is the market impact.
Then there's the spread problem. My 340% backtest used the tester's default spread handling, which on my broker's history averaged about 10 points on GBPJPY. Real spread during the London session runs 25–40 points, and triple that around news. I re-ran the same backtest with a fixed 35-point spread and the return collapsed to 41%. Same EA. Same history. The MT5 copy trading signals guide touches on why signal providers' backtests always look better than their live results — same mechanics at work.
| Modeling mode | Tick source | Speed | Trust it for |
|---|---|---|---|
| Every tick (real ticks) | Broker tick history | Slow | Scalpers, news EAs |
| Every tick | Generated from M1 | Medium | Rough screening only |
| M1 OHLC | 4 prices per minute | Fast | Daily+ strategies |
| Open prices only | Bar open only | Fastest | Optimization passes |
The Curve-Fitting Trap That Cost Me Four Months
After the live failure, I did what every stubborn trader does: I optimized. I ran the tester optimizer across 2,000 parameter combinations on 2020–2024 data and found settings that turned the 340% into 610%. I was thrilled. I forward-tested those "perfect" settings on the first half of 2025 data the optimizer had never seen. They lost 18%.
That is curve-fitting in its purest form: the optimizer didn't find a better strategy, it memorized the past. It's the difference between learning to drive and memorizing one specific road. Change the road — new market regime — and you're in a ditch.
My rule now, learned over four wasted months: optimize on 70% of history, validate on the untouched 30%. If the out-of-sample result keeps at least half the in-sample edge, the strategy might be real. If it collapses, the edge was never there. Walk-forward analysis — rolling the optimization window forward through history — is the stricter version, and it's what I run before any EA touches real money. Investopedia's backtesting guide explains the statistical trap clearly, and the MQL5 articles section has solid walk-forward tutorials.
Setting Up the Calculator Fields for Tester Validation
When I validate a backtest, the DarkWolf MT5 Signal Calculator gets these inputs — this is the pre-flight checklist before I believe any tester report.
| Input Field | What it means | Sample value |
|---|---|---|
| Account balance | The live account size you'll actually trade | $3,000 |
| Risk percent | Per-trade risk the EA's sizing implies | 2% |
| Stop loss (pips) | EA's average stop + real spread + slippage | 23 pips (15 + 8 costs) |
| Take profit (pips) | EA's average target minus spread | 27 pips (30 − 3 spread) |
If the calculator's position size differs wildly from what the EA traded in the backtest, the backtest's risk profile is fiction. Fix the EA's sizing first, re-run the tester, then compare again. Two iterations of this loop usually kill 90% of fantasy results.
$100,000 Account: Backtest vs. Forward Test, Month by Month
Here's the honest comparison from my rebuilt scalper — same EA, tested properly this time, on a $100,000 prop account. Backtest used real ticks, 35-point fixed spread, $7 commission per lot, and 20 points of slippage per trade.
| Month | Backtest return | Forward (demo) return | Trades |
|---|---|---|---|
| January | +2.1% | +1.4% | 64 |
| February | +1.6% | +1.9% | 58 |
| March | −0.8% | −1.2% | 71 |
| April | +2.8% | +2.2% | 66 |
The numbers rhyme. Not identical — forward testing never matches exactly — but the same shape: small wins, one red month, no blowups. That's what an honest backtest looks like. Boring. Believable. The 340% fantasy never rhymed with anything.
The Modeling-Quality Math
Two formulas decide whether your backtest deserves trust. The first is expectancy — the average dollars per trade after all costs. The second is my own "honesty ratio": forward result divided by backtest result.
| Variable | Meaning | Example value |
|---|---|---|
| Win rate / Avg win | From tester report, costs included | 58% / $42 |
| Loss rate / Avg loss | Including slipped stops | 42% / $38 |
| Costs | Spread + commission + slippage per trade | $6.40 |
| Honesty ratio | Above 0.5 = believable; below = fantasy | 0.72 (healthy) |
Expectancy with the example numbers: (0.58 × 42) − (0.42 × 38) − 6.40 = 24.36 − 15.96 − 6.40 = $2.00 per trade. Thin, but positive — and it survived forward testing. The fantasy backtest had an expectancy of $11 per trade. Fantasy is always generous.
My Pre-Launch Backtest Checklist
Every EA now passes this gauntlet before live money. No exceptions, no shortcuts.
One: real-ticks modeling, full history, the EA's actual magic number and inputs — no "I'll fix the inputs later." Two: spread set to the 90th percentile of what I actually see, not the average. Three: commission and swap entered in the tester inputs, because the tester doesn't know your broker's fee schedule. Four: slippage of at least 20 points per side. Five: the 70/30 in-sample/out-of-sample split. Six: a minimum of 300 trades in the test — fewer than that and you're reading noise, not signal. Seven: two weeks of forward demo on a VPS running 24/7, because the tester can't simulate disconnects, and disconnects are where EAs actually die.
The risk management settings guide pairs well with this checklist — backtest honesty means nothing if the live risk settings don't match the test. And BabyPips has a good beginner-friendly walkthrough of the tester interface if the buttons themselves are still unfamiliar.
Backtest Questions I Get Every Week
Why does my EA perform worse live than in the strategy tester?
Usually a mix of four things: generated ticks instead of real ticks, unrealistic spreads, zero slippage and commission in the test, and curve-fitted parameters. Fix all four and the gap shrinks dramatically — it never fully disappears, because live markets include disconnects and human panic that no tester simulates.
Should I use "Every tick" or "Every tick based on real ticks"?
Real ticks, always, for any strategy that trades intraday. "Every tick" synthesizes price movement inside M1 candles, which invents fills your EA would never get. The only cost is speed — real-tick tests run slower, so use them for final validation, not for thousand-pass optimization runs.
How much history do I need for a trustworthy backtest?
At least 300 trades or two full years, whichever is more. Fewer trades and you're measuring luck. Make sure the period includes different regimes — a trend year and a chop year. An EA tested only on 2020's trends will look like a genius and trade like a tourist in 2023's ranges.
Does the MT5 tester account for swap and commission?
Only if you enter them. The tester has input fields for commission and models swap from the symbol specification, but many traders leave commission at zero. On a scalper doing 60 trades a month, $7-per-lot commission is the difference between profit and loss — enter it every time.
Can the strategy tester backtest multi-currency EAs?
Yes — MT5's tester supports multi-currency testing natively, unlike MT4. Select your symbols in the tester settings and the EA can read all of them. Just remember that real-tick data must be downloaded for every symbol, or the tester silently falls back to generated ticks for the missing ones.
How long should I forward-test before going live?
Minimum two weeks on demo under live-like conditions, ideally a month. Forward testing catches what backtests can't: VPS disconnects, weekend gaps, broker feed quirks, and your own emotional reaction to the EA's drawdown. If you can't stomach the demo drawdown, you won't survive the live one.