Guide · Risk and position sizing · Updated 2026-09-19

Scaling exposure to a volatility target cut the deepest drawdown by thirteen points, left realised volatility higher than doing nothing, and added time under water

Three questions decide whether the rule is worth running, and generic treatments answer none of them: what it did to the numbers it is named after, how sensitive that is to the one parameter nobody agrees on, and what it was holding on the day before each of the big falls. Everything below is computed from the exchange session files with every parameter stated.

The asymmetry the whole rule stands on

A sizing rule that reacts to volatility is only worth writing if volatility is more forecastable than direction. It is, by a distance that can be put on a page rather than asserted.

At one session the signed move carries an autocorrelation of -0.0026, which is nothing, while the absolute move carries +0.2336. Push the separation out and the sign stays at nothing while the size decays slowly, still +0.0695 a month apart. Measured on non overlapping windows rather than lags, the asymmetry is sharper still wherever there are enough windows to measure it.

At twenty one session windows, one window's volatility accounts for 31.2 per cent of the variance in the next window's volatility, and one window's return accounts for 0.0 per cent of the next window's return. The second is zero to one decimal place, on 71 independent pairs, so no ratio between the two means anything. At 63 session windows only 20 pairs remain, too few to separate either figure from noise.

Why the asymmetry exists also tells you when it will fail. A repeatable direction is an instruction to buy or sell and gets competed away by whoever finds it first, so surviving predictability in the sign has to be small. Repeatable size is not an instruction by itself. It persists because the mechanisms producing a large move produce more large moves: information arrives in clusters, a shock forces leveraged holders to reduce and the reduction is itself a large move, option hedging demands scale with the move, and market makers widen after a loss so the same order flow moves the price further. None of that is arbitraged away, because knowing that tomorrow will be violent does not tell you which way. That is the raw material a sizing rule can use and a direction rule cannot. The stationarity of a price series is the neighbouring question: the level is not stationary while the size of its changes is closer to being so.

The rule, stated so it can be checked

For each session, take the standard deviation of the previous L daily returns of the broad benchmark index, all from sessions strictly before the one being sized, and annualise by the square root of 252. Divide a stated annual volatility target by that estimate, clip at a stated cap, hold that exposure for the next session with the remainder of capital idle, and repeat.

Parameters used throughout: target 15 per cent a year, caps of 2.0 and 1.0, and seven lookbacks of 5, 10, 21, 42, 63, 126 and 252 sessions. All seven start on 2014-01-08, so the longest lookback is burned in before the common window opens and every variant is measured on identical ground. Idle cash earns zero and levered exposure is charged nothing; both are simplifications stated rather than hidden, and both affect a return path rather than any of the four quantities reported here. Two integrity guards run before the page is written: no calendar year from 2013 to 2025 may hold fewer than 235 sessions, and no gap between consecutive sessions may exceed seven calendar days. The measured minimum is 240 sessions and the largest gap 6 days. A missing block would masquerade as one enormous return and corrupt every volatility reading near it.

The cap is reported both ways throughout, for a reason that becomes the finding. A cap of 2.0 is the rule as usually written, free to lever up when the market is quiet. A cap of 1.0 is the same rule with the levering half removed, able only to cut.

The measured frontier across lookbacks

Read the drawdown column down and then the volatility column, because they do not tell the same story. Maximum drawdown improves at every lookback, most at 21 sessions, reaching -25.91 against the unscaled -38.44. Realised volatility does the opposite: at 5 of the seven lookbacks the scaled series was more volatile than the index it was scaling, and at 5 sessions it reached 20.07 per cent against a stated target of 15. A rule named after volatility missed its own target upward by more than five points at its most reactive setting.

The shortest lookback is the instructive failure. At 5 sessions the estimate is built from a handful of returns, so it swings violently and the exposure swings with it: 41.6 annual turns to buy 3.30 percentage points of drawdown protection. The 252 session lookback bought 3.86 points for 0.97 turns, roughly 43 times cheaper. The reason sits in the false signal count.

That last column is the short lookback problem in one number. At 5 sessions the rule cut hard and reversed itself inside a month 46 times; at 21 sessions, 11 times; at 126 and 252 sessions, 0 and 0. Each round trip is two trades that produced no change in position a month later, and each was paid for.

This is the measurement that settles the lookback argument better than any of the others. A 252 session lookback produced a whole period volatility of 15.56, closest of all seven to the target, and hit the target in individual windows 42.6 per cent of the time, only 6.8 points better than the 35.8 per cent the unscaled index managed by doing nothing at all. Averaging to the right number over a decade while controlling little month to month is not volatility targeting; it is a coincidence with a rule attached. The 10 session lookback hit the target in 69.6 per cent of windows, best of the seven, with the 21 session lookback close behind at 67.6, and that is the honest case for the middle of the range.

Volatility and drawdown are separate measurements and they moved apart

Here is the finding this page exists for. At the main setting the rule improved maximum drawdown by 12.53 percentage points and made realised volatility 0.75 points worse. Those two facts are not in tension; they are what happens when a rule is allowed to lever.

Realised volatility over a decade is dominated by the ordinary body of the distribution, thousands of unremarkable sessions. Maximum drawdown is a single number set by one episode, and it lives entirely in the tail. The rule's two halves act on those two regions differently. The cutting half operates in the tail, where volatility has risen, and it is what lowers the drawdown. The levering half operates in the body, where volatility is low, and it is what raises the realised number: mean exposure at the main setting was 1.27, with more than full exposure on 74.0 per cent of sessions. Multiply a long quiet stretch by 1.27 and the everyday noise of the series goes up, and it goes up on far more sessions than the tail ever contains.

The cut only table proves the decomposition. Remove the levering half and realised volatility falls from 16.65 to 13.13, below the unscaled 15.90, while maximum drawdown improves further still to -24.01. On this record, and on these four measurements, the half of the rule that cuts did all of the useful work and the half that levers did none of it. That is not a general law and one window cannot make it one, but it is what the data says and it inverts the usual presentation, in which the levering is treated as the elegant symmetry that makes the rule complete.

A maximum drawdown is also a single draw from a distribution rather than a property of a rule, which is the load bearing caveat on the entire drawdown column above. The sampling behaviour of a maximum drawdown is the necessary companion to this page: the improvement measured here is one realisation, and the sampling error on a maximum is large enough that a reader should treat the direction as informative and the magnitude as approximate.

What the rule did through a spike

This is the honest core and it is a matter of sequence rather than opinion. Volatility rises during a decline, because a decline is made of large moves and large moves are what the estimator measures. So the input that triggers the cut is the loss. The rule cannot act before the loss that informs it, and a page that describes volatility targeting as protection without saying that is describing something that does not exist.

In the largest episode, a fall of 38.44 per cent over 48 sessions, the rule entered holding 1.27 times exposure and did not cut below full exposure until 17 sessions in. It reached 0.29 at the trough, which is real protection, and it reached it 48 sessions after the peak. The concentration is the reason the lateness costs so much: the five worst single sessions inside that episode compound to -33.78 per cent of a 38.44 per cent fall. Nearly the whole decline arrived in five sessions, and a twenty one session estimator cannot respond inside five sessions.

Look down the exposure column. At every one of the 6 largest declines in the record the rule was holding more than full exposure at the peak, averaging 1.50 against an overall average of 1.27. It was, on average, 18 per cent bigger than its own normal size going into the worst episodes of the decade.

The largest position in the record sat in the calmest conditions in the record

This is the failure mode that generic pages omit, and on this data it is not a theoretical worry. Targeting trailing volatility means levering up when volatility is low. Volatility is low in calm periods. Calm periods are where volatility spikes start, because a spike is by definition a transition from calm. The rule is therefore largest exactly where it is most exposed to being wrong.

The clearest measured instance is the episode beginning 2026-01-02. Trailing 21 session volatility at that peak was 7.23 per cent, in the calmest 5 per cent of the whole record, so the rule was pinned at its cap of 2.0, the largest position it is permitted to hold. What followed was a 15.18 per cent decline over 58 sessions, and the rule did not cut below full exposure until 31 sessions in, by which point 15 per cent of the fall had already been taken at more than full size.

Now the counterweight, because the failure mode is real and the naive version of it is wrong. Sort every session by the exposure the rule was holding and ask what the index did over the following twenty one sessions.

On average, high exposure was followed by calmer conditions, not worse ones: the sessions held at 1.50 and above saw an average worst forward move of -1.68 per cent against -2.66 per cent for the middle band. That is the rule being right, and being right on the great majority of days is why it survives casual inspection. The entire problem sits in the last column. The worst forward path attached to any bucket is -37.24 per cent, and it followed an exposure of 0.98, a whisker under full size, while the worst after exposure above full size was -31.40 per cent. The rule is not wrong on average. It is wrong in the tail, which is the only place a maximum drawdown is decided, and it walked into the worst of it at close to full size.

The same shape appears session by session. Across the window the scaled series took a larger loss than the index on 1,074 of 1,428 down sessions, 75 per cent of them, because it was carrying more than full exposure on most ordinary days. Its single worst session was 2015-08-24, at -6.38 per cent against the index's -5.92 per cent on that day, while on the index's own worst session it held 0.29 and took -3.72 per cent against -12.98 per cent. The rule cut the extreme tail roughly in half and magnified the ordinary body. Cap it at full exposure and the magnification count falls to 0. Position sizing from first principles is where that trade sits, because a sizing rule is a choice about which part of the distribution to pay attention to.

Turnover is the bill, and it is the part most often left off

Trading to the target every session at the main setting costs 10.47 annual turns. Trading only when the target differs from the held exposure by more than a fifth costs 4.29, a reduction of 59 per cent, at the price of 2.52 points of drawdown and 0.43 points of volatility. On the cut only version the same band takes turnover from 1.77 to 0.69, 61 per cent less, and moves drawdown by -0.69 points and volatility by -0.11, differences small enough to be noise on a sample this size. A band is close to free on the version of the rule that only cuts.

The arithmetic that converts turnover into a number is one line: annual drag equals one way turnover multiplied by your own one way cost per unit traded. The second term is the one nobody measures honestly, because it is not the visible charge. It is the charge plus the half spread paid plus the impact caused plus the slippage between the signal close and the execution, and on a rule that trades every session in the same direction as everyone else running the same rule, the impact term is not small. Modelling transaction costs inside a backtest and the real cost of an Indian trade are where that second number gets built. Until it is built, a turnover figure of 10.47 a year is a quantity with no price attached, and it is the reason a sizing overlay that looks clean gross can be worthless net.

Where the rule did not help, stated plainly

Four places, all measured on this window.

It did not reduce time under water anywhere. Every one of the fourteen settings tested, seven lookbacks at each of two caps, spent a larger share of sessions below its own prior peak than the unscaled index's 89.9 per cent. At the main setting the figure was 92.2 per cent. Reducing the depth of a fall reduces the rise required to recover it, but the rule is still de levered through the early recovery, because volatility stays elevated after a decline ends, and the two effects roughly cancelled here and then some. The longest single spell below a prior peak was 500 sessions unscaled and 565 sessions at the shortest lookback, which is worse by 65 sessions.

It did not hit its own target at most settings. Only 4 of seven lookbacks finished within a point and a half of the stated 15, and the one that finished closest controlled month to month volatility only 6.8 points better than doing nothing.

It did not protect against a gap and cannot. Exposure is fixed before the session opens from data ending at the previous close. On 2024-06-04 the index moved -5.93 per cent in a session after a stretch whose trailing volatility read 14.97 per cent, the rule was holding 1.002, and the scaled loss was -5.94 per cent, larger than the index's own move that day. What the rule reduces is the tail of a prolonged decline, where volatility has time to rise and the estimator has time to respond. A one session repricing is outside its reach by construction.

It did not help at the shortest lookback on any dimension. At 5 sessions the rule was worse than the unscaled index on realised volatility, worse on time under water, worse on the longest spell under water, better on maximum drawdown by 3.30 points, the least of any lookback, and it charged 41.6 annual turns for the privilege. Three of four measurements worse and the fourth improved least of all is not a marginal setting; it is a refutation of the reflex that a faster estimate is a more responsive rule.

What the measurement is actually for

Not for choosing a lookback. The frontier above has no minimum that holds across the four quantities, and any number picked off it would be picked because it looked best on this window, which is the error the whole exercise is meant to expose. A result conditioned on market state makes the same point from the other side: a single figure is an average over conditions weighted by how often each happened to occur, and a decade of Indian sessions holds far fewer independent episodes than it holds sessions.

What it is for is knowing which promise the rule keeps. It keeps a promise about the depth of a prolonged decline and it keeps that promise late. It makes no promise about the time spent recovering, and on this record it made that worse. It makes no promise at all about a single session repricing. And if it is permitted to lever, it makes the ordinary days noisier in exchange, which shows up in the measurement it is named after and not in the one people quote.

The separation that matters most is the one the cut only column keeps making. A rule that cuts when volatility rises and a rule that levers when volatility falls are two different rules with two different justifications, and running them as one because the formula is symmetrical is a choice, not a consequence. On this window, measured four ways, the first did the work and the second bought the risk. Anyone running the combined version should be able to say what the levering half is for, in their own words, with a number attached.