A strategy that works only inside one carefully chosen date range has not shown a stable edge. Backtest window sensitivity tests the same rules over shifted, equal-length historical windows to reveal how much the result depends on where the sample starts and ends.
This is not another parameter search. The strategy, costs and account model stay fixed. Only the historical window moves. If that small administrative change rewrites the verdict, the date range was doing more work than the trading rule.
What backtest window sensitivity measures
Backtest window sensitivity measures how performance changes when the test period moves while its length and every trading rule remain fixed. It is a stability check on the sample, not a method for finding the prettiest period.
The distinction matters because several tests answer different questions:
| Test | What changes | Question answered |
|---|---|---|
| Parameter sensitivity | A stop, target, threshold or other setting | Does the edge survive nearby settings? |
| Regime analysis | The market state used to group results | Where does the strategy make or lose money? |
| Walk-forward analysis | Training and subsequent unseen test segments | Does a repeated design process survive forward tests? |
| Window sensitivity | The start and end of an equal-length sample | Does one historical slice control the conclusion? |
The parameter-sensitivity guide looks for a plateau around the settings. The market-regime guide separates trend, range, calm and stress. Window sensitivity sits beside them: it asks whether the complete report is robust to a different cut of history.
Why the date range can rewrite a prop result
A shifted date range changes the sequence that the account experiences. Even when most observations overlap, the removed segment and the newly added segment can alter trades, open-position state, compounding and the route to a loss rule.
Common mechanisms include:
- Warm-up and startup state. Indicators and multi-timeframe rules may need prior bars before they are fully formed. A run that begins with different history can initialise them differently.
- A concentrated profit source. One cluster of trades may carry the whole sample. Moving the window can expose that dependence without changing the code.
- A concentrated loss source. A difficult sequence near an endpoint may disappear from one report and enter another.
- Balance-based sizing. Earlier gains or losses can change later position sizes, so identical later signals need not produce the same cash path.
- Open-position carry. A start or end boundary can include a signal but exclude the position that came before it, or truncate the trade that follows.
- Account clocks. Daily loss logic and overlapping exposure depend on order and timing. A prop result is a path through constraints, not just a pile of closed trades.
This is why a single full-period return cannot answer the prop question. The strategy may remain profitable in several windows while failing the account constraint in one of them. The funding model keeps the equity path and the account limits beside the headline result instead of treating profit as the only outcome.
How to run a clean rolling-window check in cTrader
A clean check fixes the cBot, preset, symbol, timeframe, data source, starting capital, spread, commission, swap and sizing before moving the dates. Otherwise the exercise mixes sample sensitivity with a second experiment.
cTrader's official backtesting documentation lets the tester choose an exact historical period and records the resulting operations and equity statistics. That makes the platform suitable for a controlled comparison, provided the inputs are written down with each report.
Use this protocol:
- Choose an equal window length and a fixed shift interval before inspecting outcomes.
- Run every available window under the same preset and execution assumptions.
- Save the complete report, including the date boundaries and data source.
- Compare expectancy, trade count, profit factor and the worst floating-equity drawdown.
- Record whether the applicable account rule would have been crossed, even when the final return stayed positive.
- Inspect which trades entered or left the sample when a metric changed sharply.
- Publish the range of outcomes, not only the strongest run.
Overlapping windows are useful for diagnosis, but they are not independent observations because much of their data is shared. The output should therefore be read as a sensitivity map, not as a larger sample size or a probability forecast.
The walk-forward testing explainer covers the stricter train-then-test process. Sliding a fixed backtest across history does not become walk-forward analysis merely because the dates move.
Read the pattern before reading the average
The useful result is the shape across windows, not one blended statistic. Averages can hide a strategy that alternates between strong and unusable periods.
| Pattern across windows | What it suggests | What to inspect next |
|---|---|---|
| Similar direction, different magnitude | The edge may be present but regime-dependent | Costs, exposure and adverse windows |
| Sign flips after small shifts | The evidence is fragile to sample choice | Endpoint trades, warm-up and concentrated clusters |
| Return survives but drawdown breaches appear | The edge and the prop constraint are different questions | Floating equity, overlap and daily loss clocks |
| One window dominates every summary | The headline may be event-dependent | Trade contribution and nearby date ranges |
| All windows degrade after realistic costs | The gross edge may not be executable | Spread, commission, swap and slippage assumptions |
Do not solve an ugly pattern by shortening the sample until it disappears. That turns a diagnostic into selection bias. The honest response is to state where the strategy worked, where it did not and whether the intended use requires stability that the evidence has not shown.
What this check cannot prove
Window stability cannot prove that a strategy will work live. Historical windows still come from one realised market path, overlapping runs reuse observations, and future regimes can differ from all of them.
It also does not mean every strategy must behave uniformly. A rule designed for a declared regime can vary outside that regime. The defect is not variation; the defect is presenting regime dependence as a universal edge or choosing the one period that conceals it.
Research on backtest overfitting formalises the broader danger of selecting winners after repeated trials. Window sensitivity addresses a narrower operational question: would the same fixed strategy still deserve the same description if the calendar boundary moved? It complements out-of-sample testing rather than replacing it.
How realbacktesting makes the sample visible
realbacktesting is a trading-software studio for cTrader built around results a trader can inspect and reproduce. Its published FTMO-line method identifies the cTrader server M1 data window as 2021-2026 and confirms the drawdown measure on a 30% out-of-sample hold-out.
Those are test conditions, not a live track record or a promise about a future window. The methodology page makes the data, execution, costs and hold-out visible so another trader can rerun the test and challenge the sample rather than trust a screenshot.
Frequently asked
Is changing the backtest date range curve-fitting?
It becomes curve-fitting when the date range is moved until a favourable result appears and only that run is kept. A predeclared sensitivity check does the opposite: it preserves every equal-length window to expose dependence on sample choice.
Is window sensitivity the same as walk-forward testing?
No. Window sensitivity moves a fixed strategy across historical samples. Walk-forward analysis repeatedly trains or selects on past data and evaluates the process on the next unseen segment.
Should overlapping windows be treated as separate evidence?
No. They share much of the same history, so they are correlated views of one path. Their purpose is to locate fragility, not to multiply the sample count.
Which metric matters most for a prop backtest?
No single metric is sufficient. A prop review needs the return distribution, trade contribution and floating-equity path beside the applicable daily and overall loss rules.
The stubborn takeaway
A backtest does not become robust because the date range is long. It becomes more credible when the same fixed rules survive reasonable changes to where that range begins and ends, and every failed window stays in the report.