Guide · Research method · Updated 2026-09-19

The worst decline in a record is a draw, and a longer record has a deeper one by construction

Every figure below is computed from a stated simulation or measured from a stated daily price series. None is quoted.

What the number actually is, and why that settles most of the argument

At every point in an equity curve, measure how far it sits below the highest level it has previously reached. That is the drawdown series, and it has one entry per day in the record. Maximum drawdown is a single number pulled out of it: the largest entry.

That sentence places the statistic in a specific family. It is not an average, a spread or a ratio. It is a maximum, and maxima behave unlike every other summary in common use. An average of more data is a better average. A maximum of more data is a larger maximum.

The consequence needs no assumption about markets. A maximum over ten years of a record includes every decline in the first three years of it, plus everything after. The largest member of a set cannot be smaller than the largest member of a subset. Extending a record can only hold the reported figure where it is or push it deeper. It can never pull it back. A longer record has a deeper maximum drawdown by construction.

One process, nothing changing, six record lengths

Remove every competing explanation: independent daily returns from one fixed distribution, no drift, no volatility clustering, no regime shift, nothing that could make one stretch genuinely riskier than another. Generate 20,000 runs of twenty years with a fixed seed and read the maximum drawdown at six checkpoints, in units of the process's own annual volatility.

The median goes from 1.08 annual volatilities at one year to 3.55 at ten and 5.08 at twenty, with nothing in the generating process differing between those rows. The growth has a shape, and it is the same one that governs the uncertainty of an average.

The simulated column and the square root column sit close enough that the mechanism is not in doubt. Quadruple the record and the typical worst decline roughly doubles. To put one illustrative scale on it, at an annual volatility of 15 per cent those medians are declines of about 15 per cent over a year and 41 per cent over ten.

Width is the other half, and it is the half that gets forgotten

At any single length the answer is spread over a wide range, so one observed value is a draw rather than a reading. A three year record of this process lands between 1.43 and 2.58 annual volatilities across the middle half of runs, and reaches 3.81 at the ninety fifth percentile. Same process, same length, a spread of nearly three to one, and nothing in the number itself says where in that band it fell.

So neighbouring lengths overlap. The ninetieth percentile of a three year record, 3.34, sits well above the tenth percentile of a ten year record, 2.12, and a short record beats a long one outright 14 per cent of the time.

The single run above shows the nested version. Its worst decline within three years was one number; within ten it was 3.6 times deeper, arriving after year five. Across all 20,000 runs, the worst of the first three years failed to survive as the worst of the first ten in 91 per cent of them. At one year against ten it fails 99 per cent of the time.

Comparing two records of different lengths is not a comparison

Two strategies are placed side by side, one reports a shallower worst decline, and it is read as the safer one. That reading is available only if the records are the same length, and they almost never are. A newer fund has a shorter record, a strategy rebuilt after a rule change has a shorter record, and a backtest has whatever length the data somebody happened to hold had.

The simulation prices the error. Take two runs of the identical process, track one for three years and the other for ten, and compare the figures as though they were comparable. The ten year record is judged riskier 86 per cent of the time; at one year against five, 92 per cent. There is no difference between the processes. The comparison is measuring the window, with a strong and predictable bias against whichever record is longer.

That bias runs opposite to the one most people guard against. A short record hides risk, which is understood. It also wins the comparison, so the strategy with less history looks better on the risk measure precisely because there is less of it. A track record is penalised for existing.

The repair is complete and unglamorous. Truncate the longer record to the length of the shorter, measure both over that common window, and state the window. Anything else compares sample sizes in the costume of a risk comparison. The same discipline governs what a result is benchmarked against, where the alternative is a free choice and the choice decides the answer.

The same effect, measured on the real Indian record

In real data, where declines are not independent and volatility clusters, the effect is starker. What follows is measured on the daily close of the broad Indian equity index across 3,385 trading days, 2013-01-01 to 2026-09-18, read from the exchange's own daily index file.

8 of the 13 rows report exactly 38.4 per cent, because every one of them contains the same episode. The rest report between 15.2 and 17.2 per cent. The asset is identical in every row. A fund holding this exposure since 2021 would truthfully report 17.2 per cent and one holding precisely the same thing since 2013 would truthfully report 38.4 per cent, and a reader comparing them would conclude something about the managers that is an artefact of the launch dates.

Now the spread across records of equal length: each calendar year is a fresh one year record, with the reference peak reset each January.

Across 14 such years the deepest within year decline ranges from 4.1 per cent in 2017 to 38.4 per cent in 2020, a spread of 9.4 times, median 12.0 per cent. Any one of those figures, quoted alone as the risk of holding this exposure for a year, is one draw from that row presented as a property.

The point at which the number stops being a statistic at all

Rolling the window across the record exposes something sharper.

All 926 ten year windows in this record report the same number, 38.4 per cent, to the last decimal. So do 55 per cent of the five year windows. Over a long window the maximum drawdown of this record is not a summary of the record. It is the size of one episode, the fall from 2020-01-14 to 2020-03-23, restated once for every window long enough to contain it.

That is what an extreme value statistic degenerates into on a finite record: a sample size of one. The other 3,384 daily observations contributed nothing beyond establishing the peak it fell from. More history cannot make it more reliable, because history does not add information to a maximum. It adds opportunities for the maximum to be exceeded.

Which direction the bias runs, and by how much

The statistic is used to anticipate what a holding period will feel like, so the question is whether it misleads in a direction. Simulate a record, then the period that follows it from the same unchanged process, and count how often the future is worse than the past said. Here the process carries a positive drift, set so its annualised mean is 0.5 times its annualised volatility, stated as a parameter rather than claimed about any market.

The first and last rows are the instrument checking itself: when record and holding period are the same length and drawn from the same process, the probability that the second is deeper must be exactly one half by symmetry, and the simulation returns 49.5 per cent. That row is also the best case, and it is already uncomfortable. Even when the future is exactly as long as the record, the record's worst decline is beaten half the time. It is not a ceiling or a worst case. It is a median.

Every other row is worse, because holding periods are normally longer than the records used to justify them. A three year record followed by a ten year holding period is exceeded 81 per cent of the time, at typically 1.56 times the depth the record showed. As an estimate of what is coming, the observed maximum is too shallow, and it gets worse the longer you intend to stay. That is also why a small edge is hard to hold on to at all, the subject of how much data it takes to tell an edge from luck.

A real drift does not exempt a method from this

Drift helps. The ten year median falls from 3.55 annual volatilities to 2.25 and the growth rate flattens below the square root law, visible in the right hand column of the growth table. It still grows at every step, and the comparison error survives: two identical runs of this process, tracked for three years and for ten, still produce the verdict that the longer one is riskier 81 per cent of the time.

The reason is structural. Drift shortens the expected time to recover from a decline but puts no ceiling on how deep one can go, so the maximum keeps getting fresh chances to be exceeded for as long as the record runs. A better method has a shallower distribution of drawdowns. It does not have a stationary maximum, because no method does.

What survives the record length problem

Three summaries of the drawdown series do not collapse onto a single observation.

The distribution rather than its largest member. The series has one entry per day. Its median, upper quartile and ninety fifth percentile use all of them, and each converges as the record lengthens.

Time under water, as a share of the record. Time below a prior peak is the part of a decline people actually experience, and the share form of it is stable where the maximum is not.

On the real record it sits between 86 and 89 per cent at every window length from one year to ten, and the spread across windows narrows as the window grows, which is how a usable statistic behaves. Over the same lengths the measured maximum drawdown roughly tripled. One of these numbers is about the asset and the other is about your window.

The recovery question. Depth and duration rank episodes differently. The deepest episode here fell 38.4 per cent from 2020-01-14 to 2020-03-23, taking 48 trading days down and 158 to regain the old peak. The longest stretch below a prior peak ran 500 trading days, about 2.0 years, from 2015-03-04 to 2017-03-10, and was shallower. Falling hard and recovering in months is a different experience from drifting below a peak for two years, and maximum drawdown scores the first as the worse of the two.

The absolute duration grows with the record while the share falls and steadies, the same contrast the real data shows. Wherever a quantity can be stated as a share of the record rather than a running total, it should be, because that is the form that stops the record length from doing the talking.

What India actually requires anyone to be told

By circular SEBI/HO/IMD/IMD-PoD-2/P/CIR/2025/6 of 17 January 2025, the Securities and Exchange Board of India made disclosure of a risk adjusted return measure compulsory for equity oriented mutual fund schemes for the first time. The measure is the Information Ratio, defined in the circular as the portfolio rate of return less the benchmark rate of return, divided by the standard deviation of that excess return, against the scheme's Tier 1 benchmark, with the standard deviation computed from daily return values. It must be published on the asset manager's website alongside performance every day, and the industry body must carry the same figures in a comparable, downloadable, machine readable format. The provisions came into force within three months of issuance.

The choice of statistic is the point. The risk term India mandated is a standard deviation, which uses every observation and settles down as the record lengthens, so two schemes with different histories can be compared without the histories deciding the answer. No drawdown figure appears in that circular, and none is required elsewhere in the mandated retail disclosure set. The stress test disclosures for certain fund categories report how long a portfolio would take to sell down, a liquidity exercise and not a drawdown measure either.

The two rows that rest on one observation are the two that must carry a record length attached every time they are quoted.

The tolerance you set, and the draw you get

The sequence is familiar. A method is tested over whatever history was available and reports a worst decline. The person sizes their position, or picks the product, against that figure: they could sit through that much and no more. The number has become a psychological contract.

Everything above says it was one draw from a wide distribution, over a window shorter than the one they are about to live through, and that it will be beaten 81 per cent of the time by a ten year hold. So the most likely single outcome is not that the method fails. It is that the method performs exactly as expected and the person still passes through a decline they had explicitly decided they would not tolerate, at which point the evidence in front of them is indistinguishable from evidence that the method has broken.

The repair is a change of habit. Set tolerance against the distribution of drawdowns over the horizon you intend to hold, not the single worst figure in a shorter record. Assume before committing that the deepest decline ahead is worse than the deepest already seen, because for any holding period longer than the test that is the probable case rather than the pessimistic one. And size the position so being wrong about the depth is survivable, which is the only part of this a person fully controls.

None of that needs better data or a cleverer statistic. It needs treating one number as what it is: the largest value in a list whose length somebody else chose.