Anchored vs Rolling Walk Forward Analysis: The Real Difference

Traders who build systematic forex strategies all hit the same question eventually: what is the difference between anchored and rolling trading system walk forward analysis, and which one should you trust? The short answer is that both test the same thing, which is whether optimised parameters survive on unseen data. They differ in how much history each optimisation is allowed to remember. That one design choice changes what your results actually mean.

This matters because walk forward analysis is one of the few defences against curve fitting. If you pick the wrong window structure for your strategy, you can still end up fooling yourself, only more carefully. This article explains how each method works, what each one assumes about markets, and how to choose between them. It also covers the window-sizing mistakes that quietly ruin otherwise sound tests. You get mechanics, a worked EUR/USD example, a decision framework, and a checklist you can use on your next optimisation run.

Table of Contents

• Walk Forward Analysis: The Core Mechanism
• How Anchored Walk Forward Works
• How Rolling Walk Forward Works
• The Difference Between Anchored and Rolling Walk Forward Analysis, Side by Side
• Choosing a Method for Forex Strategies
• Window Sizing and Common Pitfalls
• FAQ
• Conclusion

Walk Forward Analysis: The Core Mechanism

Walk forward analysis splits historical data into a sequence of paired segments. Each pair has an in-sample (IS) window, where parameters are optimised, and an out-of-sample (OOS) window, where those parameters are traded blind. You then step forward and repeat. Finally, you stitch the OOS segments together into a single equity curve made only of trades the optimiser never saw.

The method was formalised for trading by Robert Pardo. It is summarised well in the overview of walk forward optimisation. The logic is simple. A single backtest tells you how well parameters fit the past. Walk forward tells you how well your optimisation process generalises.

Every walk forward test needs four decisions:

• IS window length: how much data each optimisation sees.
• OOS window length: how long the chosen parameters trade before re-optimisation.
• Step size: usually equal to the OOS length, so OOS segments do not overlap.
• Window behaviour: anchored or rolling. This is the subject of this article.

Here is a point many traders miss. If the OOS segments line up, anchored and rolling tests produce OOS periods covering the same calendar dates. Only the IS windows differ. So any difference in results comes entirely from what the optimiser was allowed to learn from.

How Anchored Walk Forward Works

In an anchored walk forward, the IS window always starts at the same point, the anchor. With each step, the end of the IS window moves forward, so the window grows. The OOS window moves forward by a fixed amount each time.

Worked Example: EUR/USD, 2014–2024

1. Optimise on 2014–2017. Trade OOS in 2018.
2. Optimise on 2014–2018. Trade OOS in 2019.
3. Optimise on 2014–2019. Trade OOS in 2020.
4. Continue until the final OOS year, 2024.

By the last run, the optimiser sees ten years of data. The first run saw four.

Strengths

• More data per optimisation. Larger samples mean more trades and lower variance in fitness metrics.
• Parameter stability. Each new year adds only a small share to the IS set, so optimal parameters tend to drift slowly.
• Regime coverage. Later windows include several volatility regimes, such as the 2015 SNB shock, the 2020 crash and the 2022 rate cycle.

Weaknesses

• Slow adaptation. Old data carries the same weight as recent data. A structural change gets diluted.
• Uneven windows. Early and late runs have very different statistical power, so their results are not directly comparable.
• Heavy overlap. Consecutive IS windows share most of their data. Consecutive optimisations are therefore not independent tests.

In practice, anchored analysis asks one question: does a strategy optimised on all available history keep working?

How Rolling Walk Forward Works

In a rolling walk forward, also called a sliding or moving window, the IS window keeps a fixed length. Both its start and its end move forward each step. Old data drops out as new data comes in.

Same Data, Rolling Structure

1. Optimise on 2014–2017. Trade OOS in 2018.
2. Optimise on 2015–2018. Trade OOS in 2019.
3. Optimise on 2016–2019. Trade OOS in 2020.
4. Continue with a constant four-year IS window.

Strengths

• Adaptivity. Parameters reflect recent market behaviour, which matters when volatility or correlation structures shift.
• Comparable windows. Every optimisation uses the same amount of data, so the runs are statistically consistent.
• Diagnostic power. Parameter drift across windows becomes visible. Large jumps reveal a fragile fitness landscape.

Weaknesses

• Smaller samples. A fixed window caps the trade count. Low-frequency systems can fall below meaningful sample sizes.
• Parameter noise. The optimiser can chase short-lived conditions, which is a form of overfitting in itself.
• Regime amnesia. Rare events drop out of the window and stop shaping the parameters.

Rolling analysis asks a different question: does regular re-optimisation on recent data add value?

The Difference Between Anchored and Rolling Walk Forward Analysis, Side by Side

Underneath the mechanics, the two methods rest on opposite assumptions about markets. Anchored assumes the edge is stationary: more history means better estimates. Rolling assumes the edge is non-stationary: old history is contamination.

• IS window start: fixed (anchored) vs moving (rolling).
• IS window length: grows each step vs stays constant.
• Weight of old data: permanent vs eventually zero.
• Parameter behaviour: smooth and slow vs responsive and noisier.
• Sample size per run: increasing vs capped.
• Best suited to: structural edges vs regime-dependent edges.
• Main risk: stale parameters vs fitting to noise.

Both methods are judged with the same metric. Walk forward efficiency (WFE) is annualised OOS return divided by annualised IS return. Pardo’s commonly cited guideline treats a WFE of roughly 50–60% or higher as evidence that the optimisation is robust. Values near 100% are unusual and worth a closer look. A WFE well above 100% usually points to luck or a mistake rather than skill.

WFE is not the only check. Also look at:

• The percentage of profitable OOS windows.
• Maximum OOS drawdown compared with IS drawdown.
• Variance of optimal parameters from window to window.
• Whether OOS profit comes from one or two windows only.

A useful insight: run both. If anchored and rolling give similar OOS results, your edge does not depend on the window, which is a strong sign of robustness. If only rolling works, your strategy depends on frequent re-optimisation, and you must actually do that live.

Choosing a Method for Forex Strategies

Forex has particular features that affect the choice. Major pairs trade almost 24 hours a day, five days a week. Liquidity is deep, but behaviour changes with central bank cycles, and volatility regimes can persist for years. A strategy’s logic should therefore drive the method, not habit.

Use Anchored When

• The edge is structural, such as session-open behaviour, carry or calendar effects.
• The strategy trades rarely, for example fewer than 50 trades a year.
• You plan to re-optimise rarely, or not at all, in live trading.
• You want a conservative, low-turnover parameter set.

Use Rolling When

• The edge is volatility or regime sensitive, such as breakout thresholds tied to ATR.
• The strategy trades often, so short windows still contain hundreds of trades.
• You will re-optimise on a fixed schedule in live trading.
• You need to show that parameter drift is controlled.

Consider a Hybrid

There are practical middle routes:

• Capped anchored: grow the window up to a maximum, then roll it.
• Weighted anchored: keep all history but weight recent trades more heavily in the fitness function.
• Parameter plateau selection: in either method, pick the centre of a stable profitable region instead of the single best peak.

One firm rule applies here: the walk forward method must match the live process. If you validate with a rolling three-year window re-optimised quarterly, you must trade exactly that way. Any mismatch makes the test meaningless.

Window Sizing and Common Pitfalls

Window sizing causes more failures than the choice of method itself. Getting it wrong can make either approach useless.

Sizing Guidelines

• OOS to IS ratio: 10–30% is the usual range. A one-year OOS after a four-year IS gives 25%.
• Trade counts: aim for at least 30 trades per OOS window and well over 100 per IS window. More is better.
• Number of runs: at least 5–10 walk forward steps. Three steps is an anecdote, not evidence.
• Regime coverage: each IS window should cover at least one full volatility cycle for the pair.

Pitfalls That Invalidate Results

1. Optimising the window lengths. If you test fifteen IS/OOS combinations and report the best, you have curve-fitted the walk forward itself. Window length is a parameter too.
2. Repeated testing. Running many walk forwards and keeping the winner brings back selection bias. Bailey and López de Prado measure this directly in their work on the probability of backtest overfitting.
3. Unrealistic costs. Fixed spreads hide rollover widening and news spikes. Model variable spread, commission, slippage and swap.
4. Look-ahead leakage. Indicators that need warm-up bars must not pull data across the IS/OOS boundary in a way live trading could not.
5. Platform limitations. MetaTrader 5’s built-in forward test performs only one IS/OOS split. That is a single holdout, not a walk forward. True walk forward needs scripting or dedicated software.

Actionable Checklist

• Fix window lengths before looking at any results.
• Run both anchored and rolling and compare WFE.
• Plot optimal parameters for each window and reject erratic results.
• Judge only the stitched OOS equity curve.
• Match the live re-optimisation schedule exactly to the test.

FAQ

Is anchored or rolling walk forward analysis more reliable?

Neither is better in general. Anchored is more reliable for stationary, low-frequency edges. Rolling is more reliable for regime-sensitive, high-frequency strategies. Running both and comparing the results is the strongest test.

What is a good walk forward efficiency score?

A WFE of about 50–60% or higher is commonly treated as acceptable. Below that, the optimisation process probably fits noise. Scores well above 100% should be investigated, not celebrated.

How long should the out-of-sample window be?

Typically 10–30% of the in-sample length. It must also hold enough trades to be meaningful, ideally 30 or more. Match it to how often you plan to re-optimise live.

Can I do walk forward analysis in MetaTrader 5?

Not natively. MT5’s forward option creates a single in-sample/out-of-sample split. A full anchored or rolling walk forward needs custom scripts or third-party software that automates repeated optimisation.

Conclusion

So, what is the difference between anchored and rolling trading system walk forward analysis? Anchored keeps every bar of history and grows the optimisation window. That gives stable parameters but slow adaptation. Rolling keeps a fixed window, adapts to recent conditions, and risks chasing noise. The OOS periods can be identical. What differs is what the optimiser remembers.

Choose based on your strategy’s edge, its trade frequency, and how you will actually re-optimise live. Fix window sizes in advance. Run both methods. Trust only the stitched out-of-sample curve. Take your current EA, run both walk forward types this week, and see whether its edge survives.