The risk of a portfolio is not the average of its parts

Two holdings, and a risk that is smaller than the average

A portfolio's risk is not the average of its holdings' risks. Holdings that rise and fall on different schedules spend part of the time cancelling each other, so the combination moves less than the pieces do — while the expected return stays at the plain weighted average. That gap between the two averages is the entire subject.

Take two holdings and choose their numbers to make the arithmetic easy. The first is expected to return 12% a year and its yearly returns scatter around that by 20 percentage points. The second is expected to return 8% and scatters by 10. Put half the money in each.

That spread has a standard name — the standard deviation of the returns — and this article calls it scatter throughout, because that is what it measures. Every figure here is an illustration with its inputs stated; none of it describes any actual asset, scheme or period.

The expected return of the blend is the easy half: half of 12 plus half of 8, which is 10%. That is exact, and — this is the part worth holding on to — it stays at 10% no matter how the two holdings behave relative to each other. Nothing about their relationship changes the average of their expectations.

The variability is where it stops being an average. If the two always moved in perfect lockstep, the blend would scatter by half of 20 plus half of 10, which is 15 — the weighted average, exactly as intuition expects. Lockstep is the only case where intuition is right. The table below recomputes the same 50/50 blend for a range of assumed relationships between the two, with everything else held fixed.

Read down the last column first. It never moves. Read down the fourth and the scatter falls from 15 to 5 while the expected return sits still at 10% — and the holdings, the weights and the expectations were never touched. Only the assumed relationship changed. The variance column is an intermediate step in squared units and is not meant to be read on its own; the scatter is its square root.

The whole model lives in one term

Where does the reduction come from? From a term that has no equivalent in ordinary averaging, and it is worth seeing written out because everything later in this article is an argument about that one term.

The blend's variance is the first holding's weight squared times its variance, plus the second holding's weight squared times its variance, plus twice the two weights multiplied together, multiplied by both scatters, multiplied by the correlation between them. With the numbers above: a quarter of 400 is 100, a quarter of 100 is 25, and the third piece is 2 × 0.25 × 20 × 10 × the correlation, which is 100 times the correlation. So the variance is 125 plus 100 times whatever the correlation turns out to be — which is exactly the table.

The first two pieces are properties of each holding on its own. The third belongs to neither of them; it is a property of the pair. At its heart is how much the two move together — the covariance — and that same quantity divided by both scatters, which puts it on a fixed scale from −1 to +1, is the correlation. A portfolio of 20 holdings has 20 variance terms and 190 pairs, so almost all of the arithmetic is about relationships rather than about holdings.

Two consequences follow immediately, and they are mechanical rather than debatable. The first is that a holding has no risk contribution of its own: the same fund added to two different portfolios changes each of them by a different amount, because the pairwise terms differ. The second is that a portfolio can be made steadier without lowering its expected return at all — not by finding calmer holdings, but by finding holdings whose bad periods do not coincide. What that cancellation can and cannot remove is set out in diversification, which owns that argument.

Adding the riskier holding can make the portfolio safer

Sweep the weight of the first holding from nothing to everything, and for each weight compute the pair of numbers the model cares about: expected return and scatter. Plot them against each other and you get a curve. That curve is the efficient frontier, and the shape it takes is the model's one genuinely surprising claim.

Keep the two illustrative holdings from above and assume they are unrelated, so the cross term vanishes. Four points on the curve, all of them re-derivable from the formula in the previous section:

• All in the calmer holding: scatter 10.0, expected return 8.0%.
• One-fifth in the riskier one: scatter 8.9, expected return 8.8%.
• Half and half: scatter 11.2, expected return 10.0%.
• All in the riskier holding: scatter 20.0, expected return 12.0%.

Look at the first two. Moving a fifth of the money out of the calmer holding and into the one that scatters twice as much produced a portfolio that both scatters less and expects more. The riskier holding made it steadier. That is not a paradox and it is not clever; it is the cross term doing its work, and under these inputs it is exact arithmetic that anyone can redo in three lines.

It also gives the curve its shape. Below that one-fifth point every blend is beaten on both counts at once — more scatter, less expected return — by some blend above it. Those blends are dominated, and the model's word for the rest of the curve is efficient: nothing available offers more expected return at that level of scatter. The blend at the very bottom, where the scatter is smallest, has a name worth remembering for the next section — the minimum-variance portfolio.

So far, none of this is contestable. Given the inputs, the arithmetic follows. The rest of the article is about the inputs, and about the word risk quietly standing in for scatter throughout.

Assumption one: risk means scatter, and scatter has no direction

The model measures risk as variance — the average squared distance of a return from its own mean. Squaring is what makes the arithmetic work, because squared distances add up in the tidy way the previous sections relied on. It is also what removes the sign.

A year that came in ten points above expectation and a year that came in ten points below contribute identically to the risk number. The model has no way to prefer one. A holding whose surprises were mostly pleasant is scored as exactly as risky as one whose surprises were mostly not, provided they scattered by the same amount. No investor has ever felt that — and an optimiser, meaning any routine that searches the weights for the best return-and-scatter combination, will happily trade away a pattern of upside surprises to reduce a number that was counting them as damage.

There is a second, quieter assumption underneath the first. Summarising a whole distribution of outcomes by its middle and its spread is only sufficient if the distribution has no other interesting features — no lopsidedness, no unusually heavy extremes. Real return series have both. That matters because the frontier is drawn from the middle of the distribution while the outcomes that end a plan come from its edge, so the statistic doing the describing is the one that says least about the part you care about.

Two structures make this concrete. A strategy that collects a small premium regularly and occasionally pays out a large one produces a long run of tiny, near-identical returns, so its measured scatter is low right up until the payout. And an asset that is revalued rarely rather than priced continuously records fewer movements simply because nobody wrote them down. Neither is safer. Both score as safer, and an optimiser fed those scores rewards them for being poorly measured.

The measures that split the number by direction — downside deviation, capture ratios, the depth of the worst fall — are the subject of fund risk measures. What is worth carrying here is narrower: the frontier is a picture drawn in a unit that treats good surprises as damage, and every point on it inherits that.

Assumption two: the inputs are estimates, and the optimiser trusts them completely

The frontier needs three sets of inputs: an expected return for each holding, a scatter for each, and a correlation for every pair. None of them is observable. All of them are computed from a stretch of history and then used as though they were facts about the future.

The three are not estimated equally well, and the difference is structural rather than a matter of having better data. Under the model's own assumption that returns are independent draws from a fixed distribution, sampling more often tells you a great deal more about how much returns scatter and almost nothing more about their average — because the average is pinned down by where the series started and where it ended, and switching from monthly to daily observations does not move either. 10 years of daily data is still 10 years of evidence about a mean.

So the least reliable of the three inputs is the one the frontier is most sensitive to. And an optimiser does not merely use that input — it searches for the weights that maximise against it. Suppose two holdings genuinely have the same expected return, and the sample happens to put one at 12% and the other at 8% purely because of which decade the window covered. An optimiser told to maximise expected return for a given scatter will not split the difference. It will tilt hard towards the 12% one, because that is what the instruction says to do.

Name the failure so it is recognisable: the optimiser treats estimation error as information, and concentrates precisely where the error is largest. It gets worse as you feed it more candidates, since more candidates means more chances that one of them has a large upward error to find. The output looks precise — weights to two decimal places — and that precision is entirely a property of the inputs, not of the world.

This is the practical reason the bottom of the frontier behaves differently from the rest of it. The minimum-variance weights need no expected returns, so the input estimated worst never enters the calculation. That is a mechanical statement about which quantities appear in which formula. Whether minimum-variance portfolios have actually held up better outside the window they were fitted on is a separate, empirical question with a substantial literature behind it, and this article does not answer it.

The same trap has a milder everyday form. A scheme's past record is also an estimate from a window, and reading one is subject to the same end-point problem set out in rolling returns. Past performance describes what happened over the period measured and does not indicate future results.

Assumption three: one correlation, assumed to hold

The cross term is the whole benefit, and the cross term is proportional to the correlation. So the frontier is drawn by assuming a single number for each pair — and a correlation coefficient is not a property of two assets. It is a property of two assets over a chosen window.

What that means arithmetically is easy to show, because the earlier example can simply be recomputed. Keep the same two holdings, the same 50/50 split, the same expectations, and change only the assumed relationship from unrelated to strongly together. The variance goes from 125 to 205, so the scatter goes from 11.2 to 14.3. Nothing was bought or sold. The position was sized against the smaller figure.

Now the mechanism that makes this more than a hypothetical. A correlation computed over a stretch of ordinary trading measures how the two behaved when nothing much was driving both of them. A stretch containing a common shock measures something else, because a shock is by definition the thing that moves everything at once. The single coefficient you are handed is an average across those states, and the frontier drawn from it is an average frontier — not the one that applies on the day the diversification is needed.

Stated carefully, that is two separate claims and they deserve different weight. That a coefficient reflects the window it was computed over is mechanical, and it is the subject of reading a correlation heatmap. That correlations across real Indian assets actually rise during broad falls is empirical, and it needs a dataset and a period stated alongside it before anyone should act as though it were established. This article makes the first claim and not the second.

The mistake to recognise in yourself is narrower than "correlations change". It is accepting a diversification benefit computed on calm data and then sizing a position with it — because the size you can hold is decided by the bad state, and the number that justified the size was measured in the good one. The useful question about any published correlation is not whether it is high or low. It is what window produced it, and whether that window contained anything that frightened anyone.

Assumption four: one period, no cashflows, and no costs

The model allocates once, at the start, and looks at the result at the end. There is no second decision in it. That single-period frame quietly excludes most of what actually happens to a portfolio, and the exclusions are worth listing because each one is a real cost the frontier does not charge.

• No rebalancing — left alone, the weights drift as the holdings return differently, so the portfolio held at the end is not the one chosen from the frontier at the start. Restoring it is a decision the model never makes and rebalancing covers.
• No money in or out — the frame has no room for monthly contributions or for withdrawals, so the order in which returns arrive cannot matter inside it. For anyone drawing an income it matters more than the average does.
• No costs and no tax — the curve is drawn on gross returns. Every move along it triggers brokerage, possibly an exit load, and a capital-gains event.
• No lock-ins and no lot sizes — holdings are assumed infinitely divisible and freely tradable. An equity-linked savings scheme carries a statutory lock-in of three years running separately from each instalment, so a holding the frontier says to trim may simply not be available to trim.
• No preferences beyond two — the investor is assumed to care about exactly the mean and the scatter. Not the horizon, not the liquidity, not whether their job income falls at the same time as their portfolio.

None of these is a flaw in the arithmetic. They are the boundary of what the arithmetic was built to describe, and the failure is using the picture outside it. A frontier point is a statement about one holding period under a set of estimates — it is not a plan, and it says nothing about the horizon over which any of it is meant to be held. Choosing that split is the subject of asset allocation, where the starting question is the fall you could sit through rather than the return you would like.

What survives once the assumptions are stripped out

A change of question survives; the optimiser and its weights do not. Four assumptions, all of them known to be imperfect, and what is left is more than the frontier and less than an answer.

Before this construction, the natural thing to ask about a holding was whether it was risky. After it, the question is what the holding does to the risk of everything else already owned — and that question survives every assumption above, because answering it usefully does not require a number. It requires noticing that the same holding is a different decision for two different people.

Three readings follow from that, and none of them needs an optimiser.

1. No such thing as a risky holding, only a risky combination. Whether a volatile position is reckless depends entirely on what it sits beside, which is why comparing two schemes on their own scatter answers a question nobody asked.
2. Different names, one position. Several holdings that respond to the same driver are one position wearing several labels, and the cross terms are what expose it. That is the argument diversification makes in full.
3. Precision borrowed from the inputs. Weights carried to two decimal places from three years of monthly data are a precise summary of three years of monthly data, and of nothing else.

The trade-off in using the model at all is the same one every summary statistic makes. Compressing a portfolio into a mean and a scatter is what allows two portfolios to be compared at all; it is also what discards the direction of the surprises, the shape of the extremes and the order in which everything arrived. The comparability and the loss are the same operation, and there is no version that keeps one without the other. What the trade-off between the two summary numbers actually claims, and where it stops claiming it, is risk and return.

A model is a set of assumptions. Knowing the four above does not make the frontier wrong — it makes the frontier a conditional statement, which is what it always was. The version to distrust is the one presented without them.

Testing the sensitivity instead of arguing about it

Every claim above is arithmetic, but how far a frontier moves when the window changes is not something to reason about — it is something to compute on a price history and look at.

FNOTrader's Mutual Funds app runs on the full AMFI history of net asset value, the per-unit price of a scheme — around 34 million rows — and reports rolling-return distributions across every start date available, alongside maximum drawdown and comparison against a chosen benchmark. Recomputing the same pair of schemes over two different windows is the direct way to see how much of a diversification benefit was the window rather than the pair. The stock backtesting universe covers roughly 2,390 stocks and 17 NSE sector and size indices, so the same exercise runs on indices rather than schemes.

FNOTrader is not a SEBI-registered investment adviser or research analyst, and nothing here is a recommendation to buy, hold or avoid any scheme, security or allocation. The illustrative weights in this article are arithmetic under inputs chosen to be easy to follow, not a suggestion about what anyone should own.

Common questions

Continue reading

More in Advanced Investing · App: Mutual Funds · Definitions: glossary · Free tools: calculators · All: every article