What is forex backtesting?
Forex backtesting is running a set of trading rules over historical currency price data to estimate how those rules would have performed. You define an entry condition, an exit, a stop and a position size; the tester walks the data bar by bar and records every trade the rules would have taken, along with the profit or loss of each.
The output is a track record of trades that never happened. That is not a criticism — it is the entire point. A backtest is a cheap way to reject an idea before it costs real money, and rejection is the outcome you should expect most of the time.
The danger is treating it as a forecast rather than a rejection filter. A backtest cannot tell you a strategy will work. It can tell you a strategy did not work, which is a far more reliable and far more useful signal. Traders who internalise that distinction stop looking for the backtest that confirms the idea and start looking for the one that would have broken it.
Why do most forex backtests overstate the edge?
Most backtests overstate the edge because every ambiguous moment gets resolved favourably by default, and five specific leaks account for the overwhelming majority of the inflation.
The same-bar stop and target. When a single bar contains both your stop level and your take-profit level, the tester has to decide which was hit first — and the bar itself does not say. Many testers pick the profitable one, some pick at random, and a few simply pick the target because it makes the equity curve smoother. Over several hundred trades that one assumption can invert a result outright. The honest resolution is to look at finer-resolution data inside that bar and, where no finer data exists, book the loss.
Look-ahead bias. Using information that would not have been available at the moment of the decision. The obvious form — using a bar's close to trigger a trade at that same bar's open — is easy to spot. The subtle forms are not: indicator warm-ups that peek forward, higher-timeframe series resampled in a way that leaks the incomplete bar, and any filter computed across the entire dataset before the walk begins. Normalising your data using statistics from the whole history is look-ahead, even though it feels like preprocessing.
Survivorship in the parameter set. Testing 200 variants and reporting the best one is not a backtest, it is a search. The best of 200 random strategies looks good by construction, and nothing about the winning result tells you which of the two you found.
Costs that do not match reality. A 0.4 pip spread assumption on a pair that widens to 3 pips around news is not a small approximation — for a strategy targeting 10 pips it is a quarter of the edge, applied to every trade.
Not enough trades. Forty trades tells you almost nothing. The variance of a 40-trade sample is wide enough to swallow most real edges and to manufacture several fake ones.
How much history do you need for a forex backtest?
You need enough history to include market regimes that are unkind to your idea, which in practice means several years rather than several months. A trend-following rule tested only across a strongly trending period has never actually met a range, and a mean-reversion rule tested only through chop has never met a trend.
Bar count matters more than calendar length. A strategy trading M15 accumulates decisions roughly sixteen times faster than one trading H4, so one year of M15 can be considerably more informative than four years of H4. What you are really accumulating is independent decisions, not days on a chart.
Depth matters as much as span. Testing on daily bars a strategy you intend to trade intraday hides everything that happens inside the day, which is precisely where your stop lives. If your stop and target are both inside one daily candle, a daily-bar backtest is guessing at the single most important fact about every trade.
There is also a practical failure mode worth naming: silent truncation. You ask for ten years, the platform holds four for that instrument and timeframe, and it runs the four without saying so. The result is labelled as a decade test and describes less than half of one. Any tool that reports coverage only as a headline number can hide this, and the honest behaviour is to state the window actually tested and which end is missing.
What makes a backtest trustworthy?
A trustworthy backtest has three properties, and all three are checkable before you believe a single number.
It is genuinely out-of-sample. The rules and parameters were fixed on one period and graded on a later, unseen one. If the parameters were chosen using the same data that produced the headline result, that result is a description of the past rather than evidence about the future. This is the single largest dividing line in backtesting, and most published results fall on the wrong side of it.
It carries its sample size everywhere. Every metric should arrive with the number of trades behind it. A 78% win rate on 23 trades and a 78% win rate on 2,300 trades are entirely different claims wearing the same number, and only one of them is worth acting on.
It states what it could not resolve. Missing data, ambiguous bars, filters that never fired, instruments where coverage was thin. A tester that silently smooths over these is telling you a story rather than reporting a measurement, and the smoothing always runs in the flattering direction.
Backtesting versus forward testing: which do you need?
Backtesting is fast, cheap and rests on assumptions; forward testing — running a fixed rule set on live or simulated data going forward — is slow, expensive and rests on far fewer. You need both, in that order, and for different reasons.
The honest sequence is backtest to reject, forward test to confirm. Backtesting lets you discard ninety ideas in an afternoon. Forward testing lets you check that the one survivor still behaves when nobody can accidentally peek at the future, when spreads are whatever the broker says they are that morning, and when your fills happen at real prices rather than assumed ones.
A strategy that survives a decade out-of-sample and then falls apart in three months of paper trading was probably fitted to something in the history you did not notice — or the backtest was modelling execution more kindly than reality does. Both are worth finding out before the position sizes grow.
Neither approach removes the question of how many trades it takes before a result means anything. For most retail-scale edges that number is in the hundreds, which is exactly why forward testing alone is such a slow way to learn anything: at ten trades a month, a 300-trade confirmation takes two and a half years.
How do you start backtesting a forex strategy?
Start by writing the strategy down precisely enough that it could be wrong. This is the actual work, and it is worth doing even if you never run a single test — most strategies dissolve the moment someone tries to state them exactly.
"Buy when the trend is up and there is a pullback" is not testable. "Buy when price closes above the 50-period EMA, then pulls back to touch it without closing below, with a stop at the prior swing low and a target at twice the risk" is. The second version can be disproved. The first can only be argued about.
Then choose the market and timeframe you actually intend to trade, pick a window long enough to include unfavourable conditions, set your costs conservatively, and run it. Read the out-of-sample section rather than the headline. If the tool does not separate in-sample from out-of-sample, treat every number in it as provisional.
Finally, write down what result would make you abandon the idea before you look at the output. Deciding the threshold afterwards is how a marginal result becomes a promising one.