Forex Backtesting: How to Test a Strategy Without Lookahead

Backtesting is a reconstruction: take a fixed rule, walk forward through available observations, and record decisions, assumed fills and results. For a forex strategy, that means defining the currency pair, price data, position size, entries, exits and transaction costs. The question is what those rules produced under that model, not whether attractive trades can be found afterward.

A manual test can do this with a chart replay and a ledger. An automated test delegates the same sequence to software. Automation makes repetition easier; it does not repair a rule that quietly uses future prices or a fill assumption the available data cannot support.

Imagine checking only whether price eventually reached a target. That leaves unanswered whether a stop was reached first, which side of the quote was tradable, and whether the signal existed at entry. A credible test records these decisions before interpreting its final number.

Keep three questions separate: what the rule observed, what the model assumed could execute, and what the resulting ledger contains.

Write a short specification that another person could follow without asking which trades you liked. “Buy when momentum improves” is not enough. State which prices define momentum, when they become available, and what happens when conditions conflict. Use a dated version so an amended rule does not silently replace the one originally tested.

• Instrument and observations: symbol, timeframe, trading hours, time zone and permitted missing-data treatment.
• Entry and exit: exact conditions, signal timing, eligible fill event and position state.
• Sizing: fixed units or a fully specified sizing rule, including whether existing positions can be increased.
• Execution and costs: price side, spread, commission, slippage and any overnight financing.
• Boundaries: indicator warm-up, treatment of open positions and signals at the final observation.

The replay uses a deliberately simple, long-only rule. After each completed five-minute bar, compare its close with the simple average of the latest three closes, including that just-closed bar. If flat and close is above SMA3, queue a buy for the next bar’s open. If long and close is below SMA3, queue an exit for the next open. Equality changes nothing. A long position stays one position when another above-average close arrives.

This is an above/below rule, not a crossover requirement and not a strategy recommendation. Its position is always 10,000 EUR. Real research may use different rules, but changing several of them after seeing results creates a new experiment. The trading plan template helps separate the current specification from proposed revisions.

Start with the question the data can answer. An OHLC bar records open, high, low and close; those four values do not reveal every intermediate price or the order in which the high and low occurred. If both a stop and a target lie inside the same bar, the bar alone usually cannot decide which came first. Here midpoint means (Bid + Ask) / 2 at a given instant. SMA3 is the average of the midpoint closing prices of three consecutive bars.

The MetaTrader 5 documentation on ticks explains why chart prices and execution prices must be distinguished: buys and sells use different quote sides, while available charts need not display both. For a price-sensitive rule, that distinction can change whether an order qualifies and what it earns.

“Real ticks” also deserves a data check. The MQL5 testing documentation describes circumstances in which the tester generates ticks when real tick records are absent or inconsistent. Record the selected mode and investigate coverage warnings; a mode label alone does not describe every observation used.

Our replay contains 36 fictional EURUSD midpoint bars, five minutes apart, within one invented intraday session. They are not downloaded market history. There are no stop, limit or target orders to resolve inside a bar. The example teaches event order and cost accounting; it cannot tell you whether an SMA rule works on EURUSD.

Advance the replay one bar at a time. At a bar’s open, any eligible instruction queued after the previous close executes using that open and the chosen costs. Only afterward does the model observe the current bar’s close and calculate a new signal. No trade can use a close that has not yet occurred.

For example, a buy signal formed after bar 3 closes may fill at bar 4’s open. It cannot receive bar 3’s opening price. A gap between the two bars therefore matters. TradingView’s strategy documentation likewise distinguishes the creation of an order from the later tick on which its broker emulator can fill it under the default timing.

Compare the full window with the first or second half. Each selected window starts flat. Earlier closes may warm up the average, but they do not import a position, an earlier fee or a queued order. These are separate teaching runs: their results must not be added together and presented as one continuous account.

At the selected window’s end, a newly generated signal has no fill if its next open is outside that window. An existing position remains visible and is marked separately; the model does not invent an exit to make every trade look complete. Changing the window can therefore change both the trade count and the final open state.

The two halves are not an out-of-sample validation. Both belong to one small fictional example, and resetting positions changes the experiment.

For this midpoint-based model, a buy fills at midpoint open plus half the spread plus adverse slippage. An exit sell fills at midpoint open minus half the spread minus adverse slippage. Commission is a separate USD charge for each execution of the fixed 10,000-EUR position. In this EURUSD example, one pip is 0.0001 USD per EUR.

The midpoint movement alone would produce $20.00. Do not subtract spread and slippage from $18.40 again: they are already inside the fills. With the same midpoint at entry and exit, these assumptions produce −$2.30. Adding one pip to the spread reduces this round trip by $1.00, because each side receives half of the extra spread.

If a position remains open, the replay marks it at the latest modeled Bid and includes the entry commission already paid. It does not charge a future exit commission or exit slippage before an exit occurs. The displayed total is modeled P/L, not a complete account-equity calculation.

Constant costs isolate their effect; they are not a promise about actual quotes. A real test may need changing spreads, different commissions and financing across rollover. This intraday example excludes overnight swap. Read the slippage guide for the distinction between a requested price, an available quote and an actual fill.

Decide which historical observations may influence development before inspecting the evaluation period. A later period can challenge a rule that was chosen earlier, but only if its results remain outside the selection process. Repeatedly viewing the holdout and adjusting the strategy turns that period into development material.

The Forward setting in MetaTrader 5 uses a later part of the selected historical interval. Despite its name, that is not the same as waiting for future market data. Keep the platform label and the research method explicit in your notes.

Chronology also applies to indicator preparation. Earlier observations can provide the history needed to calculate an indicator at the start of an evaluation period. They should not leak a future close into the calculation or quietly create a position that the experiment was defined to start without.

If a holdout exposes a problem, document it. A revised strategy may be worth a new experiment, but the old holdout is now known data. Do not preserve the “unseen test” label simply by changing a file name or selecting another performance metric.

A favorable result is a reason to inspect the ledger, not to stop checking it. Choose several entries and exits, including an awkward boundary, and reconstruct what information was available at each decision. The most informative row may be an unfilled signal or an unattractive trade that a discretionary review would overlook.

• Lookahead: a decision uses a later close, a future-confirmed pattern or a value that was not available at the stated time.
• Overfitting: many settings are tried and only the winner is retained without recording the search.
• Selection bias: favorable dates, instruments or trades are kept while inconvenient observations disappear.
• Execution mismatch: bar data are asked to prove an intrabar fill sequence they do not contain.
• Accounting errors: repeated signals add unintended positions, costs are charged twice, or an open trade vanishes at the boundary.

TradingView’s lookahead examples show how access to future information can produce implausible historical results. A practical question is whether the same signal could have been recorded while observations arrived one at a time. If later data change an earlier decision, investigate the dependency before trusting the performance report.

Vary assumptions deliberately, not until the result looks acceptable. For example, compare the same fixed rule under several stated costs. If a small cost change reverses the result, report that sensitivity. It is information about the experiment, not a reason to hide the more demanding scenario.

Download the experiment protocol from the replay after selecting a window and costs. Keep it with the displayed scenario and ledger. Its fictional-data label matters: sharing the numbers without that context would turn a teaching exercise into a misleading performance claim.

For your own research, a protocol should let you return a month later and understand what was actually tested. A screenshot of the final profit is insufficient when the data, signal timing or cost basis has changed.

• Record the rule version, intended question and any changes made since the previous run.
• Identify the dataset, price basis, time zone, interval, omissions and preparation steps.
• Write down signal timing, fill assumptions, sizing, costs and treatment of the final open position.
• Separate development dates from evaluation dates and record whether evaluation results have already influenced choices.
• Save trade counts, the execution ledger, open-state details and failed checks alongside the summary.

The trading journal template can organize individual observations, while the protocol describes the experiment that produced them. Preserve a disappointing run as carefully as an attractive one; otherwise your research history itself becomes a selected sample.

Begin with counts: observations available, completed trades, executions charged and positions still open. A percentage without its denominator can make a tiny sample look more persuasive than it is. A queued terminal signal is not a completed trade, and an open position is not a closed win or loss.

Next reconcile closed net results with open marked results and fees already paid. Inspect where losses occur and how dependent the outcome is on a few trades. The replay contains no deposit, margin model or withdrawals, so its P/L path should not be read as a broker account balance.

Use the profit factor guide, equity curve simulator and Calmar ratio guide for the definitions and limitations of those measures. A favorable ratio cannot repair lookahead or an unrealistic fill. Nor does a few hours of fictional observations justify an annualized return or a claim about long-term drawdown.

A useful conclusion describes the conditions: which rule was tested, on what data, under which assumptions, and what remains uncertain. “This version lost money when costs increased” is a clearer research result than an unsupported promise that the next version will be profitable.

Can I backtest a forex strategy manually?

Yes. Reveal observations in chronological order, apply written rules, and record every eligible trade and assumption. Avoid moving through a finished chart and selecting only recognizable setups. Manual work needs the same cost and boundary rules as an automated test.

Should I use MetaTrader 5 or TradingView?

Choose according to the rule, data and execution behavior you need to examine. Read the relevant tester or broker-emulator settings and inspect individual fills. Different defaults can produce different results without either platform demonstrating that a strategy is profitable.

Are real ticks always necessary?

The necessary detail depends on the rule. A strategy sensitive to events within a bar needs evidence those bars may not provide. Tick data can add detail, but you still need to check quote coverage, price sides and execution assumptions. More rows do not eliminate model limitations.

How many trades make a backtest reliable?

There is no universal number that validates every strategy. Frequency, dependence between trades, market conditions and the number of variants tried all affect interpretation. Report the sample and its limitations. The small fictional replay here is deliberately unsuitable for a statistical performance claim.

Why can live results differ from a backtest?

Prices, timing, spread, available liquidity, fees and order handling may differ from the model. A test also cannot contain market conditions that have not occurred in its data. Historical results remain conditional evidence; forward observation and actual execution should be recorded separately.

Does a profitable backtest prove an edge?

No. First rule out information leakage and accounting errors, then assess how the strategy was selected and tested. A result can be internally consistent and still depend on a small or unrepresentative period. Backtesting supports an investigation; it does not guarantee a future outcome.