- What diversification actually removes
- Why the number of holdings stops mattering
- 20 stocks from one sector is one bet with 20 tickers
- Several schemes can be one portfolio
- The floor moves, and it moves the wrong way
- What diversification costs you
- The questions the count cannot answer
- Where this sits in the app
- Common questions
What diversification actually removes
Diversification removes the risk that belongs to one company — the fire, the fraud, the failed product, the order from a regulator. It cannot remove the risk that reaches every company on the same morning. That split is a mathematical property, not a rule of thumb, and everything else in this article falls out of it.
Start with why the first kind cancels. A plant burns down. A founder is arrested. A patent case goes the wrong way and a product line closes. Each of those happens to one company, on its own timetable, for reasons that have nothing to do with the company next to it in your portfolio — so when it happens to one holding, it is by construction not happening to the others. Hold enough of them and the bad surprises and the good ones land in the same period and partly offset. The work is being done by independence, not by the number of names.
Now take a change in the policy rate, or a global funding shock, or a currency move that reprices imported inputs across the whole market. That does not arrive at one holding. It arrives at all of them, in the same direction, on the same day. There is nothing left for it to offset against, because it is the same term appearing once in every holding rather than a different term appearing in each.
Put numbers on it, with the sizes chosen to be legible rather than typical. Split each holding's return into a shared part and an own part, and give both the same size — one unit each, where a unit is how far that part typically lands from its own average. For a single holding, the two combine to a spread of √(1+1), about 1.41 units. Spread the money equally over several holdings and the shared part is untouched at one unit, while the own parts average out: the average of n independent one-unit parts has a spread of 1÷√n. So the portfolio's spread is √(1 + 1/n).
Run it. Four holdings: 1.12 units. Ten: 1.05. Twenty: 1.02. A hundred: 1.005. The removable piece falls as 1/n in variance terms, so ten holdings have removed nine-tenths of it and 20 have removed 95% — and every holding after that is competing for a sliver. The floor is one unit, and one divided by 1.41 is about 0.71. Seven-tenths of the original spread survives any number of holdings you can name.
Those two parts have names, and the names arrive last because the idea is what matters. The part that cancels is idiosyncratic, or company-specific, risk. The part that survives is systematic, or market, risk. What compensation an investor is owed for bearing the second is a separate question, taken up in risk and return.
Why the number of holdings stops mattering
The section above hides an assumption, and the assumption is where almost all real portfolios come apart: it treated the own parts as independent of each other. Drop that, and the arithmetic changes shape entirely.
Take n holdings of roughly equal size, each with the same spread, and let the average correlation between any two of them be some number ρ between 0 and 1. The portfolio's variance is then the sum of two terms: 1/n, which is each holding's own variance being averaged away, and ρ(1 − 1/n), which is what they share. Written out, that is 1/n + ρ(1 − 1/n). As n grows, the first term vanishes and the second converges on ρ. The portfolio's spread converges on √ρ times a single holding's.
Read that limit again, because it is the whole argument in one line. The count enters only through 1/n, which is nearly exhausted by the time you have a couple of dozen names. What is left is ρ. The floor is set by correlation, and the count merely decides how quickly you arrive at it. The equal-split illustration above is simply the case ρ = 0.5, and √0.5 is 0.71 — the same seven-tenths.
Two chosen values show what that costs in practice. At ρ = 0.6, a 20-holding portfolio has a spread of √0.62, about 0.79 of a single holding's; going to 200 holdings takes it to √0.602, about 0.78. Ten times the names, one percentage point. At ρ = 0.2 the same move takes 0.49 to 0.45, which is √0.2 to two decimals — the floor, reached and not improvable. Identical arithmetic on the count, completely different outcome, and the difference is entirely in a number that has nothing to do with how many things you own.
The named failure mode here is the count fallacy. Counting holdings answers the half of the formula that finishes first and ignores the half that binds. A portfolio of 40 names whose owner cannot say what those names have in common has optimised the term that had already gone to nearly zero.
One honest caveat about ρ itself, which the rest of the article develops: it is not a property of your portfolio the way the number of holdings is. It is a measurement over a chosen stretch of past days, and it moves. What that measurement can and cannot support is the subject of reading a correlation heatmap, and it is worth reading before leaning on any single value of it.
20 stocks from one sector is one bet with 20 tickers
The clearest way to see ρ doing its work is a portfolio that passes every count-based test and fails the only one that matters.
Hold 20 lenders. Each has its own management, its own book, its own idiosyncratic risk — and the averaging described above works on exactly that part, as advertised. But the credit cycle, the policy rate, the regulator's next circular and the direction of asset quality are not 20 separate events. They are one event that reaches all 20 positions. The own parts averaged away; a sector-wide part sat underneath them the whole time and never entered the averaging at all. Three layers, not two: what is specific to a company, what is shared by an industry, and what is shared by the market.
The same structure produces several mistakes that look nothing alike:
- The index you already own. A market-capitalisation weighted index carries whatever sector mix its largest constituents happen to have, and that mix is a consequence of prices rather than a design. An investor who holds a broad index fund and then adds individual names from the sectors that already dominate it has concentrated, not diversified. The current weights are published by the index provider and change with prices — read them off the factsheet rather than from memory.
- One group, several listings. Companies under a common promoter group can look like unrelated businesses on a screen and still share funding lines, reputation and governance. A single event at the parent reaches every one of them, which is precisely the shared term that averaging cannot touch.
- The employer stack. Salary, stock granted by the employer, and holdings in the employer's sector are one bet held three ways. What makes this the most expensive version is the timing: the downturn that hits the shares is the same downturn that puts the income at risk, so the diversification fails at the exact moment the cash was needed.
- Small companies, one exit. A business with one plant, one product line and a handful of customers has more of its outcome riding on its own circumstances, which is the argument for holding several rather than one. What they also share is a dependence on someone being willing to take the other side — and that dependence is not specific to any one of them, so it sits in the part averaging never reaches.
A concentrated position is not a mistake in itself — it is a bet that what you know about one business is worth more than the offsetting the other 19 would have provided. That is a coherent position with a stated cost. The mistake this section names is different: it is holding a concentrated bet while believing, on the evidence of a count, that you hold a spread one. A fund built to do this deliberately says so in its category, which is the subject of sectoral and thematic funds.
Several schemes can be one portfolio
The fund version of the same error is harder to see, because the thing being counted has a different name on each statement.
SEBI defines 11 equity scheme categories, and a scheme's category constrains where it may invest. Two schemes in the same category are therefore fishing in overlapping water before either manager has made a single decision. Add that fund managers read the same disclosures and face the same liquidity limits on how large a position they can take in a smaller company, and the overlap between two same-category portfolios is structural rather than coincidental. Two schemes, one exposure, at two expense ratios.
What settles it is not the number of schemes but the weight they share. Take the two published portfolios, and for every company that appears in both, add the smaller of the two weights. That sum is the share of your money that is the same holding wearing two names. It is dull arithmetic and it takes an afternoon, and it is the only version of the question that has an answer — which is why the disclosure exists, and how to read one is covered separately.
Three cases where the answer is known before you start:
- Two trackers, one index. Funds tracking the same index hold the same portfolio by design, so holding both is one position with two folio numbers and two costs. What actually differs between trackers is expense and how closely each follows the index, which is the subject of index funds and ETFs.
- The fund-of-funds illusion. A portfolio of several schemes chosen for variety at the scheme level can hold one set of companies underneath. Counting managers is counting the layer above the exposure.
- Three brokers, one portfolio. Splitting holdings across accounts changes where they sit, not what they are exposed to. Operational spread is a real and separate thing — it is not risk diversification, and treating it as such is the cleanest example of counting the wrong noun.
Overlap has a cost with a name attached. Two expense ratios are certain, annual and compounding, while the diversification being bought with the second one is whatever the overlap arithmetic leaves once the shared weight is taken out. The first of those is known before you start. The second has to be computed. Where the genuine spread comes from — across asset classes rather than across schemes — is asset allocation, and keeping it from drifting back into concentration as prices move is rebalancing.
The floor moves, and it moves the wrong way
Everything above treated ρ as a fixed property. It is not. It is a measurement over a window of past days, and this section is about what happens to a portfolio when the number underneath it changes without a single trade being placed.
The mechanism is arithmetic and can be stated flatly. Suppose each holding's return is again a shared part plus an own part, with the shared part having variance c and each own part variance 1. The correlation between any two holdings is then c ÷ (c + 1). At c = 0.5 that is 0.33. At c = 2 it is 0.67. At c = 5 it is 0.83. Nothing about the companies changed across those three cases. Exactly one term got bigger, and the measured correlation between every pair rose because the shared part became a larger share of what was moving.
Feed that back into the floor. The spread a large portfolio converges on is √ρ times a single holding's, so a ρ that moves from 0.33 to 0.67 takes the floor from 0.57 to 0.82 of a single holding — on a portfolio whose holdings, weights and count are unchanged. The diversification did not fail. It was always a function of a quantity that is not constant, and the reader who counted names never saw that quantity at all.
The named failure mode here is the moving floor. The conditions that make a single factor dominate every holding's return are the same conditions in which the offsetting was supposed to help. A measure of protection that weakens in proportion to how much it is needed is not a flaw in the arithmetic — it is what the arithmetic says, and it is the part that gets left out of the phrase about diversification being the only free lunch available.
Two things are worth separating carefully here, because this is where financial writing usually cheats. That a measured correlation rises when the shared term grows relative to the own terms is mechanical: it follows from the definition, and the three numbers above can be re-derived by anyone. Whether that actually happened in any particular past market fall is an empirical question about a specific period and a specific set of holdings, and it needs a stated dataset and window to answer. This article makes the first claim and does not make the second, and the difference is not pedantry — a mechanism tells you what to look for, while an unattributed historical assertion tells you what to expect, which nobody is entitled to.
There is one more route by which unrelated holdings acquire a shared term, and it runs through the holder rather than the asset. An investor facing a margin call sells what can be sold rather than what they would prefer to sell. That imposes a common seller on positions with no common driver, and it does so on a timetable set by someone else's balance sheet. The same-day appearance of every panel turning one colour is described in what a risk-off day looks like. How much of it is forced selling in any given case is, again, an empirical question and not one this article answers.
What diversification costs you
An article that presents an option without its cost has not finished, and this one has two costs to state.
The first is symmetry. The averaging that stops one company's fire from mattering is the same averaging that stops one company's extraordinary decade from mattering. Both extremes are idiosyncratic; the arithmetic cannot tell them apart, and it does not try. What is being bought is a narrower distribution, which means giving up the right-hand tail to be rid of the left. Anyone who wants the transforming outcome has to accept the concentrated bet that can produce it, which is a legitimate choice and a different one.
The second is that spreading is not free of effort. Every additional holding is another set of disclosures to read, another position to size, another line to rebalance, and another set of costs to pay on the way in and out. Past the point where the 1/n term has been used up, those costs are certain and the benefit being bought with them is a sliver. The portfolio does not become dangerous; it becomes an expensive approximation of an index, held by someone doing far more work than an index fund requires.
There is a third cost that is really a warning. Adding holdings you do not understand adds new company-specific risks rather than removing existing ones, and the whole averaging argument assumed those risks were independent — an assumption you cannot check on a business you have not read. Spreading is not a substitute for knowing what you own. It changes what one mistake costs. It does not change how many mistakes you make. What the summary numbers describing a portfolio's spread can and cannot capture about any of this is the subject of the risk measures article.
The questions the count cannot answer
Everything above reduces to one substitution: stop asking how many, and start asking what is shared. The table sets out the cases in which a portfolio looks spread and is not.
| What it looks like | What is actually shared | What decides it |
|---|---|---|
| 20 companies, one industry | The industry's own driver — its cycle, its input cost, its regulator | Whether the own parts are still independent once that driver is taken out of each |
| Five equity schemes | An overlapping investable universe, often the same large positions | The weight two published portfolios share, not the number of schemes held |
| Two funds tracking one index | The identical portfolio, by design | Whether the benchmarks differ at all; if not, it is one holding |
| Several listings, one promoter group | Group funding, governance and reputation | Whether one event at the parent reaches all of them at once |
| Salary, granted stock, sector holdings | One employer and one industry cycle | Whether the holding would fall in the same event that ends the income |
| Holdings split across three brokers | Nothing — the accounts differ, the exposures do not | What is held. Never where it is held |
Notice what the last column never contains: a number to clear. There is no count that makes a portfolio diversified and no correlation reading that certifies one, because the answer depends on what the holdings share and that is specific to the holdings. A threshold here would be a recommendation wearing arithmetic, and this library does not issue those.
What the arithmetic does support is a way of reading your own portfolio. For each holding, ask what would have to go wrong for it to fall a long way, and then ask how many of the other holdings would fall for that same reason. Holdings that answer with the same event are one position. The count of genuinely different answers is closer to the thing you meant when you counted names, and where the two counts come apart is exactly where counting was misleading you.
Where this sits in the app
The parts of this that can be checked rather than argued about are the shared exposures: what a portfolio's holdings have in common by sector, and how tightly the things you hold have been moving with the things that drive them.
FNOTrader's Market Pulse Stocks Scanner covers the NSE universe with market breadth and sector strength views, so a list of holdings can be read against the sectors they sit in rather than as a set of independent names. The Options Analytics macro page carries the cross-asset correlation panel described in the correlation article — Pearson correlation of Nifty's daily returns against ten fixed drivers, over a window of 30, 60 or 90 observations, each window shown separately rather than blended. Reading the same pair across all three windows is the closest available check on whether a ρ you are relying on is a property of the relationship or of the sample.
Neither of those computes a diversification score, and there is no number in either app that certifies a portfolio as spread. That is not an omission. The floor is set by what the holdings share, the sharing is specific to those holdings, and the reader is the only person who knows what else they are exposed to — including the income that pays for all of it.
FNOTrader is not a SEBI-registered investment adviser or research analyst. Nothing here is a recommendation about any security, sector or portfolio, and nothing here is a forecast — it is the arithmetic of one property and the limits that come with it, so the decision stays with the reader.
Common questions
What risk does diversification actually remove?
The part of a holding's return that is specific to that company — a plant fire, a fraud, a failed product, an adverse order. Those events are independent across companies, so across enough holdings the bad ones and the good ones partly offset. What it cannot remove is anything that reaches every holding at once, such as a policy rate change or a global funding shock, because that is the same term appearing in each holding rather than a different term appearing in each.
How many stocks do I need to be diversified?
Fewer than the count-based rules imply, and after that the count stops being what decides it. The count contributes through a 1/n term, so each additional holding does less than the one before and the term is nearly exhausted early. What the portfolio converges on is set by the average correlation between the holdings, which the count cannot tell you — the same 20 names can leave you close to the floor or a long way from it depending on what they share. Any specific number offered as an answer is a threshold standing in for the part of the question that actually matters.
Is holding 20 stocks from one sector diversified?
Against company risk, yes; against the sector, not at all. The company-specific part is being averaged away, so that half works, but an industry has its own shared driver — its cycle, its input costs, its regulator — that sits underneath all 20 positions and never enters the averaging. There are three layers, not two: what is specific to a company, what is shared by an industry, and what is shared by the market. A count-based check sees only the first.
If I hold five equity schemes, am I diversified?
Not necessarily, and the number of schemes cannot settle it. Schemes in the same category are constrained to overlapping ground before any manager decides anything, so the same companies can appear in several portfolios at similar weights. The check is to take the two published portfolios, and for every company in both, add the smaller of the two weights. That sum is the share of your money that is one holding wearing two names — and you are paying two expense ratios for it.
Why do correlations rise in a falling market?
Because the shared part of each holding's return grows while the own parts stay where they were. If a return is a shared part plus an own part, the correlation between any two holdings is the shared variance divided by the total — so growing the numerator and leaving the rest alone raises the measured correlation by definition, from 0.33 to 0.67 to 0.83 across chosen values. Whether that has happened in any particular past episode is a separate question that needs a stated dataset and period, and no general claim about it is made here.
Can diversification protect a portfolio in a market fall?
It removes the risk of one holding being the problem, which is a real and complete protection against that specific thing. It does nothing about a fall that reaches every holding, because that is the term the averaging cannot reach. Worse, the floor it converges on is set by the average correlation, and that correlation rises when one factor dominates every holding's return — so the protection is weakest in exactly the conditions that produced the question.
Does diversification reduce returns?
It narrows the distribution at both ends. The averaging that keeps one company's disaster from mattering also keeps one company's extraordinary run from mattering, because the arithmetic cannot distinguish the two — both are company-specific. That is the trade being made: the right-hand tail is given up to be rid of the left. Whether that is worth it is a judgement about what the money is for, and it is not one that can be made for someone else.
Is splitting holdings across several brokers or demat accounts diversification?
No. Where a holding sits changes nothing about what it is exposed to, and a portfolio held in three accounts has the same shared drivers as the same portfolio held in one. Spreading across providers is a real thing and answers an operational question about access and single points of failure. It is a different question from the one this article is about, and treating it as an answer here is the cleanest case of counting the wrong noun.
Do my employer's shares count as one holding or several?
Salary, stock granted by the employer, and holdings in the employer's industry are one exposure held three ways, and the timing is what makes it expensive: the downturn that hits the shares is the same downturn that puts the income at risk. A useful test on any portfolio is to ask, for each holding, what would have to go wrong for it to fall a long way — and then count how many other holdings, and how much of the income, answer with the same event.
Continue reading
More in Stocks & Screening · App: Market Pulse Stocks Scanner · Definitions: glossary · Free tools: calculators · All: every article