Data study

Best forex EA: what 13.7 million backtests say about picking one

13 min read

Every “best forex EA” list ranks robots by their backtest. We wanted to know whether that ranking means anything, so we pulled together six of our own studies — 13.7 million backtests across forex majors, crosses, gold and indices, each one on real recorded market bars with the spread charged inside the simulation — and asked two questions: does picking the top of a backtest leaderboard work, and what do the strategies that kept working actually have in common?

What $10,000 became: the top picks, before and after they were pickedEURUSD · top 10: $106,087 when picked, $4,815 afterwards; EURUSD · top 100: $44,056 when picked, $5,294 afterwards; EURUSD · all 8,645 badged: $12,170 when picked, $8,850 afterwards; GBPUSD · top 10: $73,848 when picked, $3,650 afterwards; GBPUSD · top 1,000: $25,306 when picked, $6,393 afterwards; Gold's top 10 → GBPUSD: $99,672 when picked, $3,010 afterwardsSELECTIONWHAT $10,000 BECAME (LOG SCALE)PICKED ONAFTER$1k$3k$10k$30k$100kstarting balanceEURUSD · top 10: +961% on the selection years, −52% afterwardsEURUSD · top 10picked 2018–2023, then 2024–2026$106k$4.8k+961%−52%EURUSD · top 100: +341% on the selection years, −47% afterwardsEURUSD · top 100picked 2018–2023, then 2024–2026$44k$5.3k+341%−47%EURUSD · all 8,645 badged: +22% on the selection years, −11% afterwardsEURUSD · all 8,645 badgedpicked 2018–2023, then 2024–2026$12k$8.9k+22%−11%GBPUSD · top 10: +638% on the selection years, −63% afterwardsGBPUSD · top 10picked 2018–2023, then 2024–2025$74k$3.7k+638%−63%GBPUSD · top 1,000: +153% on the selection years, −36% afterwardsGBPUSD · top 1,000picked 2018–2023, then 2024–2025$25k$6.4k+153%−36%Gold's top 10 → GBPUSD: +897% on the selection years, −70% afterwardsGold's top 10 → GBPUSDpicked on gold 2018–2023, cable 2024–2025$100k$3.0k+897%−70%
Median compounded result of each group, on a log scale. The green dot is the record the strategies were picked on; the red dot is what the same strategies did on the years after the pick. EURUSD from our walk-forward on 8,645 badged candidates (2026 is a partial year); GBPUSD from the cable holdout; the last row was picked on gold and scored on cable. The broader the group, the smaller the fall — and every group still fell.

The short answer

There is no best forex EA you can identify from its backtest, and the top of a leaderboard is the worst place to look. On EURUSD, the ten best strategies picked on 2018–2023 had turned $10,000 into a median $106,087 (+961%); over 2024–2026 the same ten compounded to a median −52%. The same collapse shows up on GBPUSD and on gold strategies moved to GBPUSD. What held up across years has a recognisable profile instead: a slow timeframe (H4), a few dozen trades a year rather than hundreds, 1–2% risk per trade, and modest compounded returns with a worst drawdown around −22%. Judge an EA by that profile and by an honest per-year record, not by its rank.

The honest answer

Is there a best forex EA?

Not in the way the question is usually meant. “Best” assumes a forex expert advisor is a product you can rank, like a laptop: test them all, sort by return, buy the one at the top. Our data says the sort itself is the problem. A backtest ranking mostly measures which strategies happened to fit a particular stretch of history on a particular instrument, and that fit does not carry forward.

That does not mean systematic trading is hopeless. Across these studies a small minority of strategies did keep their shape across years, and they were not random: they share a timeframe, a trade frequency, a risk level and a drawdown profile. That profile is the most useful thing we can tell you, because you can check any EA against it — ours, a free one, or one you are about to pay for.

This page is the hub over our per-instrument and per-dimension studies. Each section below summarises one finding and links to the full write-up, where the method, the tables and the caveats live.

Sources

Six studies, one question

These were run at different times, on different populations, for different reasons, which is exactly why their agreement means something. Every run used our own C++ engine on real recorded bars, with spread, commission and swap charged as each trade happened.

StudyBacktestsWhat it measuredFull write-up
Timeframe study33,1002,050 EURUSD strategies, M5 to H4, each year separately, with costs on and offBest timeframe
Instrument study73,250The same strategies on all 25 instruments, 2018–2025Best pairs
Robots study55,265A 25-instrument matrix plus two selection holdouts (by instrument and by year)Do robots work?
Gold study186,6374,961 XAUUSD strategies, 8 full years, three cost modelsBest gold EA
EURUSD discovery8,098,335A strategy search plus an unfiltered pool, scored walk-forward on years after the pickBest EURUSD EA
GBPUSD transplant5,298,129215,727 gold-selected strategies re-run on GBPUSDBest GBPUSD EA
13,744,716 backtests in total. Counts are read from each study's own artifact; none of the six populations overlaps another's runs.

The core finding

Why the top of a forex EA leaderboard is the wrong place to look

The chart at the top of this page is the single most important result we hold. Three different designs, two instruments, the same outcome: the higher a group ranked on the years it was picked on, the further it fell afterwards. On EURUSD the top 10 went from +961% to −52%, the top 100 from +341% to −47%, and the whole badged pool — far less impressive on paper at +22% — lost only 11%. The ordering is upside down.

The reason is not fraud and not a bug. A ranking by in-sample return sorts by two things at once: genuine edge, which is small, and luck, which in a large population is large. The further up the ranking you go, the more of the result is luck — and luck does not repeat. GBPUSD shows the same pattern from cable’s own history: its top 10 picked on 2018–2023 compounded +638% there and −63% over 2024–2025, when the median strategy in the eligible pool did −11%.

Moving the instrument is the same test

The other way to check a ranking is to keep the years and change the market. In the robots study the top 25 strategies by in-sample return went from +15.25% on their home instrument to −18.27% across 12 others (H1). In the GBPUSD transplant the 1,000 best strategies by gold’s compounded return were profitable on cable 10.3% of the time — against 12.5% for the whole pool. Being a top gold strategy was, if anything, a slight handicap.

What this looks like in the app

A leaderboard sorted by return for a single year is the same ranking in miniature. Here is our app’s leaderboard for one account’s saved EURUSD H4 strategies in 2025, sorted by return:

RoboticEA leaderboard for EURUSD H4 in 2025 with real spread, sorted by return: 632 strategies, 595 profitable, the top row at plus 575 percent, row 13 grid_216 number 1 at plus 70.3 percent and row 15 Thistle EMA Trend Glide at plus 68.8 percent
One year, one pair, sorted by return: 595 of the 632 saved strategies were profitable in 2025, and the top row made +575.0%. Nothing on this screen says what any of them did in the other years.

Rows 13 and 15 both made about 70% in 2025. Open their cards, which re-run every year separately, and the ranking falls apart:

The strategy cards for grid_216 number 1 and Thistle EMA Trend Glide, both EURUSD H4: the first at minus 20 percent a year with 2 of 7 years profitable and a minus 77 percent worst drawdown, the second at plus 7.8 percent a year with 3 of 7 profitable and a minus 71 percent worst drawdown
The same two strategies, every year. grid_216 #1 (row 13): 2025 +70%, but −67% in 2020 and −70% in 2021 — 2 of 7 years profitable, −20%/yr, worst drawdown −77%. Thistle EMA Trend Glide (row 15): 2025 +69%, but −58% in 2023 — 3 of 7 profitable, worst drawdown −71%. The strategy diagrams are cut out of both cards (dashed line) so the results line up.

8 of 8

of the highest-averaging gold strategies compounded to a wiped-out account

Ranked by the arithmetic mean of their yearly returns, the top 8 gold strategies look spectacular — the leader averaged +368% a year. Compounded over the 8 full years, it went to −100%. One enormous year inflates a mean; it cannot rescue a balance that hit zero first.

That is the second ranking trap, and it is why every return on this page is compounded. An “average annual return” on an EA sales page can be arithmetically true and describe an account that no longer exists. Ask what the balance did over the whole period. The full analysis of both traps is in the EURUSD walk-forward and the GBPUSD transplant.

The profile

What a forex EA that holds up looks like

If ranking by return does not find a good EA, what does the small set of strategies that kept working have in common? Four things recur across the studies, and none of them is the entry signal.

H4 / H1

the only timeframes any EURUSD survivor came from

67

median trades a year among those survivors

≤ 2%

risk per trade of every survivor that states one

−22%

their median worst drawdown

1. A slow timeframe — usually H4

Of 2,050 EURUSD strategies tested on each of 8 years, 0.2% of M5 strategy-years were profitable against 37.9% on H4. Only 13 strategies made money in six or more of the 8 years, and of the 11 that also traded at least 20 times every year, 6 were H4 and the rest H1. On gold, 6 of the 8 most consistent strategies were H4. On GBPUSD, 16.4% of H4 strategies were profitable over eight years against 6.4% of H1. And in the EURUSD walk-forward, badged H4 strategies were profitable after the pick 37.8% of the time, H1 18.5%.

Daily bars are defensible but slow to judge: 74.7% of D1 strategies could not reach 20 trades in a single year, so a D1 record takes many years to mean anything. The why, and the full table, is in the best timeframe for expert advisors.

2. Tens of trades a year, not hundreds

Timeframe matters mostly because of what it does to trade count, and trade count matters because the spread is charged on every trade. On EURUSD the median M5 strategy traded 1,075 times a year and costs took 44.1% of the account a year; the median H4 strategy traded 35 times and costs took 1.6%. Switch costs off and 29.5% of M5 strategy-years flip to a profit they never earned. A scalping EA is not necessarily wrong about the market — it is paying a toll on hundreds of trades a year that a slower EA pays a few dozen times.

The floor matters too: a strategy with five trades a year has no record at all, however good it looks. That is why our app will not issue a verdict on a year with fewer than 20 trades.

3. Low risk per trade, and a drawdown you could sit through

Risk per trade does not add edge; it widens the outcome in both directions. The EURUSD walk-forward tested the identical 8,635 strategies at three risk levels and scored them on 2024–2026:

Risk per tradeMedian, picked yearsMedian, years afterProfitable afterMedian worst-year DD after
2%+14.4%−6.0%34.7%−16.6%
3%+18.6%−9.9%32.8%−24.1%
4%+21.1%−14.3%31.1%−31.1%
The same strategies at three risk settings, picked on 2018–2023 and scored on 2024–2026. More risk looks better on the selection years and worse on the years that follow.

The survivors bear this out: every EURUSD survivor that states a risk setting used 1–2% per trade, with a median worst drawdown of −21.8%. An EA advertising a big return at 5% or 10% risk per trade is showing you the top of a very wide distribution, not an edge.

4. An instrument that is cheap to trade — but no “best pair”

Across 25 instruments the share of profitable strategy-years ranged from 11.7% (FX cross) to 24.0% (index). No class came close to half. The bigger difference was cost: friction took 5.8% of the account a year on US30 and 22.3% on AUDNZD. But a “best pair” ranking does not transfer either — among H1 and H4 strategies the rank correlation between two instruments was a median ρ = 0.40, and the strategies that were best on an instrument went from +2.6% at home to −5.6% elsewhere. More in the best forex pairs for expert advisors.

Two H4 strategy cards, one above the other: a USDJPY strategy at plus 7.7 percent a year with 5 of 6 counted years profitable and a minus 22 percent worst drawdown, and a gold strategy at plus 64 percent a year whose 2025 chip reads plus 419 percent, with a minus 63 percent worst drawdown
The profile on two real cards, both H4, different markets. Top, USDJPY: +7.7%/yr, 5 of 6 counted years profitable, worst year −20%, worst drawdown −22% — dull, and the shape that held up. Bottom, gold: +64%/yr and a STRONG badge, but that average is carried by one chip — 2025 at +419%. Multiply out the other six chips and they compound to roughly −3%; the drawdown along the way was −63%. Ranked by return, the bottom card wins every time. (Strategy diagrams cut out at the dashed line.)

Check a strategy against the profile

RoboticEA re-runs every strategy on each year separately with real recorded spread, shows trades, drawdown and the losing years, and exports to MetaTrader 5. Free account, no card.

Try it free

Expectations

What good looks like, compounded

Once you stop ranking by the top of the distribution, the realistic numbers are much smaller — and much more honest. On gold, the most consistent strategies we found compounded as follows over 8 full years with real spread:

StrategyTFCompoundedYears upWorst DDTrades/yr
Trend ConfirmationH4+31.6%6 / 8−15.4%33
OrderflowH4+56.6%6 / 8−22.8%44
Volatility SqueezeH1+59.3%6 / 8−29.1%109
Multi-StepM30+60.7%6 / 8−46.5%57
CycleH4+13.5%5 / 8−17.7%33
RegimeH4+13.7%5 / 8−22.3%48
OrderflowH4+42.3%5 / 8−19.8%43
OscillatorH4+26.1%5 / 8−14.7%40
The 8 most consistent XAUUSD strategies at 3% risk, ranked by profitable years, not return. 8 of 8 finished positive. This is in-sample selection — chosen on the same years they are scored on — so read the shape, not the rows.

Compare the best compounded result in the whole gold study: +175% over 8 years, which is a real and large number — achieved while losing money in 4 of the 8 years and drawing down −57.8%. That is the honest shape of a high-return EA: lumpy, and very hard to hold through. The full story, including what gold’s spread does to it, is in the best gold expert advisor for MT5.

So when an EA claims triple-digit returns every year with a single-digit drawdown, the question is not whether it is the best EA. It is which of the traps above produced the number.

By market

The best expert advisor by market: where to read next

  • Gold (XAUUSD). A median recorded spread of 3.8 pips, charged on every trade. With costs off, 24.0% of gold strategy-years flip to a profit they did not earn. Best gold EA for MT5 →
  • EURUSD. The most-traded pair, and the cleanest walk-forward test we have: the top 10 picks fell from +961% to −52%. Best EURUSD expert advisor →
  • GBPUSD. Strategies that worked on gold, moved to cable: the median one compounded −38% over eight years. Best GBPUSD expert advisor →
  • Choosing a pair at all. All 25 instruments, what each charges a strategy, and why a league table of pairs does not travel. Best forex pairs for EAs →
  • Choosing a timeframe. M5 to D1, with costs on and off — the biggest single lever in this whole page. Best timeframe for EAs →
  • Whether any of it works. The population-level view, the cost experiment and both selection holdouts. Do forex robots work? →

Practical

How to judge any MT5 EA before you run it

You do not need our software for this. Whether the forex expert advisor is free, bought, or built yourself, these questions are ordered so the cheapest eliminations come first.

  • Was it picked on the years it is advertised with? If the backtest window is also the selection window, you are looking at the left-hand dots in the chart above. Ask for results on years — or instruments — it was not chosen on.
  • Is the return compounded? An average of yearly returns can belong to an account that went to zero: 8 of the 8 highest-averaging gold strategies did.
  • What spread and commission were charged? If the answer is not a number, stop. Turning costs off flipped 24.0% of gold strategy-years into profit.
  • Show me every year separately, losers included. A per-year record cannot be window-shopped; one compounded curve can.
  • At least 20 trades in every year? Below that, a good year is as likely to be chance as skill.
  • What timeframe, and how many trades a year? Hundreds of trades a year on M5 or M15 means paying the spread hundreds of times. Every EURUSD survivor was H4 or H1.
  • What risk per trade, and what drawdown? A return without its drawdown is half a number, and a return at high risk is the tail of a wide distribution. The profile that held up was 1–2% risk with a worst drawdown around −22%.
  • Can you read it? A compiled .ex5 you cannot inspect asks you to trust the seller’s backtest and nothing else. Prefer MQL5 source, or at least visible inputs — and re-run it yourself in the MT5 Strategy Tester on real ticks (why real ticks matter).

An EA that answers all eight may still lose money — every figure here is a backtest. But one that answers none is not a strategy; it is a screenshot.

A trade replay with gold candles, a long entry and close marked on the chart, and the account equity curve below with the maximum drawdown window shaded
Checking an EA's trades one by one: a replay of a single backtest year, with each entry and exit on the candles and the drawdown window shaded on the equity curve below. Watching what a strategy actually does is the fastest way to find out whether its record is a few lucky trades.

Limits

What this does not tell you

  • Everything here is simulation. We hold no live trading record. A strategy that fits the profile can still lose money from the day it goes live.
  • The profile is descriptive, not a recipe. H4, few trades and low risk describe what survived; they do not make a strategy profitable. Most H4 strategy-years lost money too — the median H4 strategy-year on EURUSD was −1.8%.
  • The survivor lists are in-sample. The EURUSD survivors and the consistent gold strategies were chosen on the years they are scored on. The out-of-sample evidence is the walk-forward and transplant results, and those say selection is hard.
  • Holdouts disagree with each other sometimes. In the robots study 3 of 4 time-holdout cells went negative, and one split of four pointed the opposite way to the rest. One split is not a holdout.
  • Selection methods beyond ranking are untested at this scale. We have shown that ranking by in-sample return fails. We have not shown that nothing works.

FAQ

Questions, answered

What is the best forex EA?

There is no forex EA you can identify as the best from its backtest. In our EURUSD walk-forward, the ten best strategies picked on 2018–2023 had compounded a median +961%; over 2024–2026 the same ten compounded to a median −52%. The same collapse showed up on GBPUSD and on gold strategies moved to GBPUSD. What held up across years had a recognisable profile instead: a slow timeframe (H4), a few dozen trades a year, 1–2% risk per trade and a worst drawdown around −22%. Judge any EA against that profile and an honest per-year record, not its rank.

Why do top-ranked expert advisors fail after you buy them?

Because a ranking by backtest return sorts by edge and luck at once, and the higher you go the more of the result is luck. In our data the higher a group ranked on the years it was picked on, the further it fell afterwards. Moving the instrument has the same effect: the 1,000 best strategies by gold's compounded return were profitable on GBPUSD 10.3% of the time, against 12.5% for the whole pool.

Is an EA's average annual return a good way to compare EAs?

No. An average of yearly returns can belong to an account that went to zero. Ranked by the mean of their yearly returns, all 8 of the 8 highest-averaging gold strategies in our study compounded to a wiped-out account, including one that averaged +368% a year. Always ask for the compounded result over the whole period, year by year.

What is the best timeframe for a forex EA?

Slower timeframes held up far better, mainly because they trade less and so pay the spread less often. On EURUSD, 37.9% of H4 strategy-years were profitable against 0.2% on M5, and every EURUSD strategy that was profitable in six or more of eight years while trading at least 20 times a year was H4 or H1.

How do I check an MT5 EA before running it?

Ask whether it was tested on years or instruments it was not picked on; whether the return is compounded; what spread and commission were charged (turning costs off flipped 24.0% of gold strategy-years into profit); for every year separately, losers included; for at least 20 trades in every year; for the timeframe, risk per trade and maximum drawdown; and whether you can read the code. Then re-run it yourself in the MT5 Strategy Tester on real ticks.

Every number above is computed from the six study artifacts in this repository, not typed in. The engine behind them has been run head-to-head against MetaTrader 5’s own Strategy Tester — the parity study — and here is how the workflow fits together.

Risk disclaimer

Every figure on this page is the output of a historical simulation, not a live trading record. Past performance does not predict future results. Backtests contain no information about the future and a strategy that was profitable over a historical window can lose money from the day it is deployed. Trading leveraged instruments carries a substantial risk of loss. Nothing here is investment advice or a recommendation to trade. RoboticEA is research and engineering software: it does not execute trades, hold funds or manage accounts. See our Terms §4 for the full statement.