Guide · Research method · Updated 2026-09-19

A walk-forward efficiency of 1.0 is most easily produced by a weak in-sample result

Every figure below was computed in the build script for this page, from the exchange file set on disk or a procedure run there. Where a figure comes from a simulation it is called one, with its parameters stated.

What the procedure does, and what it is protecting against

Optimising a strategy on a stretch of history and reporting how it did on that stretch is not a test. It is a description of the search. The winning parameter won partly because it suited the period and partly because something had to win, and from inside the period those components cannot be separated.

A holdout answers that by hiding a slice from the optimiser and scoring on it afterwards. It works once, and it answers a narrow question: whether one fit survived one stretch. Score a trend rule on a calm year and it looks broken. Score it on a trending year and it looks inspired.

Walk-forward replaces the single split with a sequence. Fit on a training window, score on the window immediately after it, advance both by a step, fit again on fresh data, score again on the next unseen stretch. The scored windows are joined end to end into one record.

Two things improve. The scored record is longer than a single holdout gives, and it spans several market conditions. More importantly the object under test changes. A holdout tests a parameter set. A walk-forward tests a procedure, asking whether re-fitting on recent history produces parameters that work in the period after. That is what a live system does.

What it does not do is make the result proof. The optimiser never saw the test windows, but the researcher does, and the researcher is inside the loop. Inspect a walk-forward record, then change the strategy family, the cost model or the window lengths, and that record has entered the fitting. Walk-forward stops a parameter set memorising a period. It cannot stop a person running the procedure repeatedly and keeping the run they liked.

The ratio, and the number it divides by

Walk-forward efficiency compresses the exercise into one figure: out-of-sample performance divided by in-sample performance, both annualised so windows of different lengths compare. Around 1.0 means the forward result matched the fitted one. Below about 0.5 means most of what the optimiser found did not survive. Above 1.0 means the system did better forward than on the data it was tuned to.

That last reading is where the trouble starts. It is stated as though the only way to raise the ratio is to improve the numerator. The denominator is equally available.

Read the last column down. The same forward performance is a failure at the top and remarkable at the bottom, and nothing about it changed. What changed is how much the optimiser found, and the ratio rewards finding less.

This is not a hypothetical route to a good number. It is the easiest one, and several ordinary research choices take it unintentionally. Narrowing the parameter grid lowers the best in-sample result. Adding costs lowers it. Choosing a training window over which the strategy family works poorly lowers it. Each is defensible, and each raises efficiency while making the system no better.

The degenerate case deserves naming. As the in-sample result approaches zero the ratio approaches infinity, and past zero the sign flips: a system that lost money in training and lost money forward reports a positive efficiency, because a negative divided by a negative is positive. In the control runs below, 1 per cent of folds were negative in both windows, and every one passes on the sign.

A walk-forward on real Indian data, run here

The series used below is an equal-weighted basket of the 200 most heavily traded stocks in the exchange cash file, over 1,164 trading days from 2022-01-03 to 2026-09-18. It returned 18.5 per cent annualised at 16.7 per cent volatility. Daily returns use each stock's close against its own adjusted previous close, so splits do not appear as one-day collapses. One caveat belongs in the text, not a footnote: the basket requires a full quote history, so it holds only names traded throughout. That is a survivorship filter, and it makes the basket kinder than the market was.

The strategy family is deliberately dull: long when a fast moving average of the basket sits above a slow one, flat otherwise, with 72 combinations forming the search grid. Costs of 10 basis points one way are charged on every change of position. The signal read at one close sets the next day's position, so no day is traded on information from inside itself. The base configuration trains on 250 days, tests on the 60 after it, and steps 60, giving 13 folds.

Averaged across folds the in-sample result was 26.0 per cent annualised and the out-of-sample result 15.8 per cent, an aggregate efficiency of 0.61. Before reading that number, notice how far the ratio moved fold to fold, and that the lengths the optimiser selected changed between folds too. A single aggregate figure hides both.

The window triple is three more parameters

Walk-forward is presented as the cure for parameter search, and it introduces three parameters of its own: how long to train, how long to test, how far to step. Nothing in the method fixes them. They are chosen by the researcher, the way any other parameter is chosen, by trying combinations and keeping the best answer.

Across 96 configurations the efficiency ranged from -0.21 to 0.79, with a median of 0.49. The best configuration reports roughly 1.6 times the median. The strategy did not change, the data did not change, the parameter grid did not change. Only the shape of the windows changed, and that shape is a free choice nobody is required to disclose.

That is the reimportation. The procedure was adopted to stop a researcher selecting a parameter set that suited a period, and it hands back fresh parameters that can be selected to suit a period. The second search is harder to see because it looks like a methodological choice, and those are rarely reported with the alternatives that were dropped. The only defence is procedural: fix the lengths before any result is seen, justify them by the horizon the system trades, and write that down first.

Anchored or rolling, and when each is the wrong choice

A rolling window keeps a fixed length and discards the oldest data as it advances. An anchored window fixes its left edge at the start of the history and grows. Both are legitimate. They encode different beliefs, and each is wrong in a specific way.

An anchored window assumes the relationship is stable and more history is therefore better. Where that holds it is the stronger design: the fit rests on a larger sample and is less easily pushed around by one unusual quarter. Where it fails, the window never forgets. A rule change from two years ago stays in the training set permanently, dragging the fit toward a market that no longer exists.

A rolling window assumes the opposite, and its failure is the mirror image. A short window may not hold enough observations for a small effect to be visible at all, so the optimiser fits noise and the walk-forward faithfully reports that the noise did not persist. It also discards the rare periods that show how a system behaves under stress.

Neither choice is settled by running both and keeping the better number, which is the fourth free parameter hiding behind the other three. It is settled by a view on whether the thing being measured is stable, and that comes from the market's institutional history rather than the backtest.

What a strategy with no edge scores on this test

A measurement with no control is a number, not a finding. The control here is what the same procedure reports when there is definitely nothing to find.

So take the same real daily returns, keep every one, and destroy their order by shuffling. The distribution is untouched: same mean, same variance, same fat tails, same worst day. What is gone is the sequence, the only thing a trend rule can exploit. Repeating the shuffle 500 times, fixed seed, same windows, gives the distribution of what the ratio produces from nothing.

Read the left column first, because it is what most published walk-forwards report. A single fold of a system that cannot have an edge scored above 1.0 in 35 per cent of cases and above 0.7 in 41 per cent, with a 95th percentile of 6.2. Nothing was there to be found.

The right column shows the only real defence, which is folds. Averaging 13 of them pulls the distribution in hard: the middle ninety per cent of outcomes narrows from a span of 8.6 to one of 0.87, and the share clearing 1.0 drops to 3 per cent. That is the genuine value of walk-forward, and it is entirely a sample size effect, which is what a short history takes away.

The middle column is the Indian case rather than a hypothetical. Restricting the same procedure to the most recent 620 trading days, about as close to one structural regime as this market allows, leaves 6 folds, and the share of no-edge runs clearing 1.0 goes back up to 16 per cent. Shortening the history to avoid a rule change and lengthening it to gain power are the same dial turned in opposite directions.

The mechanism is visible in the same runs. The ratio and its divisor move against each other.

The correlation between the in-sample result and the fold ratio across those 6,297 folds is -0.22. Folds where the optimiser found least reported the highest efficiency, by a factor of about 9 between the weakest fifth and the strongest. The underlying data did not differ: it was the same shuffled noise throughout. This is the asymmetry of the opening section, measured rather than argued.

Set the real run against that. The aggregate efficiency on the unshuffled basket was 0.61, about the 62nd percentile of what the shuffled series produces. It is inside the range, not outside it. That does not establish the strategy is worthless. It establishes that this test, on this much data, did not distinguish it from a system with no exploitable structure.

The ratio never mentions what you could have done instead

One more thing the figure leaves out, and it may be the largest. Walk-forward efficiency compares a system to itself. It says nothing about whether it beat the thing it was trading.

The basket itself returned 18.5 per cent annualised over this window. The walk-forward system returned 15.8 per cent across its out-of-sample windows while holding a position on 78 per cent of those days, and being long that share of the time and doing nothing else would have earned about 23.3 per cent over the same windows. A reader given only the efficiency figure of 0.61 recovers none of that, and the ratio would read the same had the basket fallen.

This also explains how the control was built. Shuffling keeps the drift. A long-or-flat rule on a drifting series earns roughly the drift multiplied by the share of time it is long, with no skill involved. Much of what efficiency credits to a trend system on a rising market is that arithmetic, and the ratio cannot separate the two because it never looks at the instrument.

The market moved underneath the window

Everything above is general. What is specific to India is that the object being walked forward over has been rebuilt repeatedly and recently, so a window long enough to fit is one that crosses a change in the rules it is fitted under.

The most recent change is the one least likely to appear in anything written before this month, and it touches the most used input in retail research. Since 3 August 2026, the closing price of an Indian stock that has derivatives contracts on it is no longer a volume-weighted average of the last thirty minutes of continuous trading. Continuous trading for those stocks stops at 3:15 pm, a closing auction runs to 3:35 pm, and the price struck there is the close. Stocks without derivatives kept the old method and still close on the 3:00 pm to 3:30 pm average. The auction reference price is the average over 3:00 pm to 3:15 pm, with a band of plus or minus 3 per cent.

Read that against a backtest. The liquid universe people research is very nearly the derivatives-eligible list, so its daily closing series is generated by one mechanism up to 31 July 2026 and another afterwards, and the split falls inside the out-of-sample window of any walk-forward run to the present. How the expiry-day settlement price should be derived from that close was then reopened: a September 2026 consultation put two options out for comment, with comments open until 3 October 2026. The mechanism is new and not yet settled.

It is fair to ask whether the change shows in the data rather than only in the circular. Weakly, it does. Take each stock's close against its own full-day volume-weighted average price, and measure how widely that gap spreads across a group of stocks each day. In the 34 trading days from 2026-08-03 to 2026-09-18 it averaged 66.6 basis points for the high-turnover group, against 60.9 in the matched 34 days from 2026-06-15 to 2026-07-31, a rise of 9.2 per cent. The mid-turnover control, which did not move to the auction, went from 82.7 to 84.0 basis points, a change of 1.6 per cent. The difference is 4.3 basis points, in the direction the mechanism predicts.

That result belongs with its weakness attached, and the weakness is this article's subject. Relabelling the same 68 days at random 20,000 times, a shift at least as large as the observed one appears with probability 0.12. That is not a settled finding, and group membership here is a turnover proxy rather than the published eligibility list. The measure moved as the rule change implies, by an amount this much data cannot separate from chance.

The arithmetic is unforgiving. One fold of the configuration above consumes 310 trading days. The expiry weekday has been on its current footing for 260 trading days and the closing auction for 34. There is no way to fit one fold entirely inside the current structure, let alone the several folds a walk-forward needs, so every walk-forward on this market trains on one regime and tests partly in another. The method cannot fix that. The problem is not in the method, it is in the data.

None of which argues against studying a market that changes. It argues for saying which regime a result belongs to, and treating a window that straddles a break as two things averaged together. A result reported without its regime is not wrong so much as unfalsifiable.

What walk-forward is actually for

Walk-forward does one thing reliably: it makes a particular self-deception expensive. A parameter set that memorised a period will not survive being re-fitted and re-scored a dozen times over, and finding that out costs an afternoon rather than a year of live capital. As an elimination tool it is efficient. What it cannot do is certify. The scored record is short, the effects are small, the window lengths are chosen rather than given, the researcher sees every result, and here the ground keeps moving.

So the figure is worth computing and not worth reporting alone. Report the two numbers it came from in absolute terms, so a reader can see whether a high ratio came from a strong forward result or a weak training one. Report how many parameter combinations were searched, and how many window configurations were tried before the one shown. Report the cost assumption. Report the null, because a ratio with nothing to compare it against is decoration. Publish all five and the claim can be checked. Publish the ratio alone and it cannot, and the two look identical.

The judgement that remains is about mechanism: why an effect should exist, what conditions it depends on, whether those still hold after the rules changed, and what would have to be true for it to stop. That is not a fallback for when the statistics run out. Given the sample sizes and regime lengths this market offers it is the larger part of the work, and treating it as a discipline with its own standards is what separates a research process from a search for a number that looks good.