Portfolio construction is the problem of deciding how much to hold of each thing. It matters more than picking the things, because the risk of a portfolio is not the average risk of its holdings. Combine assets that do not move together and the combination is less risky than its parts, without giving up the average return.
That single fact is the only free lunch in investing, and the rest of this article is about how to use it and where the standard method for using it breaks down.
Where diversification actually comes from
Two assets, each with its own expected return and volatility. Mix them and the portfolio’s expected return is just the weighted average of the two. Its volatility is not.
Volatility depends on how the assets move relative to each other. If they are less than perfectly correlated, some of their movements offset, and the combined volatility comes in below the weighted average. The lower the correlation, the larger the effect.
Notice the minimum-variance point. Adding a small amount of the riskier asset to an all-defensive portfolio reduces total risk while raising expected return. That is deeply unintuitive and it is the clearest demonstration that risk is a property of the combination, not of the components.
The efficient frontier, and what sits on it
The frontier is the set of portfolios giving the highest expected return for each level of risk. Everything below it is dominated. Two points on it get special names: the minimum-variance portfolio at the far left, and the tangency portfolio, which has the best ratio of excess return to volatility.
The framework comes from Harry Markowitz in 1952 and it remains the foundation of institutional asset allocation. It is also, applied naively, a reliable way to build a portfolio no sensible person would hold.
How many holdings is enough
A common question with a clearer answer than most.
Total risk splits into two parts. Idiosyncratic risk is specific to a company: a factory fire, a regulatory action, a fraud. Systematic risk belongs to the market as a whole. Diversification removes the first and cannot touch the second.
The removal happens quickly and then stops. Going from one holding to five cuts idiosyncratic risk dramatically. Five to twenty cuts it substantially again. Beyond roughly twenty to thirty reasonably distinct names, the marginal benefit is small, because what remains is market risk that no amount of adding stocks will diversify away.
This has two consequences people tend to get backwards. Holding four stocks is genuinely risky in a way that is avoidable and unrewarded, since nobody pays you for bearing a risk you could have removed for free. And holding two hundred stocks is not meaningfully safer than holding forty, it is just more expensive to run and harder to understand.
The caveat is the word distinct. Thirty stocks drawn from the same sector are not thirty holdings for this purpose. What matters is the number of genuinely different exposures, not the number of line items, which is the same trap as owning six funds that hold the same companies.
Why the optimiser gives answers you should not use
Feed historical returns and covariances into a mean-variance optimiser and it will hand you something concentrated, extreme and unstable. Ninety percent in one asset, nothing in several others, and a completely different answer next month from barely changed inputs.
This is not a bug in the optimiser. It is doing exactly what you asked. The problem is what you fed it.
Expected returns are almost impossible to estimate. Volatility can be estimated reasonably well from historical data because it is persistent. Expected return cannot. The historical average return of an asset over ten years carries enormous statistical uncertainty, and the optimiser treats your noisy estimate as a precise fact.
The optimiser amplifies estimation error rather than absorbing it. It is built to find extremes, so it systematically loads up on whatever asset your data happened to overstate, and avoids whatever it happened to understate. It is, memorably, an error-maximiser.
The covariance matrix is unstable too. With more assets than you have clean history for, some estimated relationships are essentially noise, and the optimiser will happily construct a hedge out of a correlation that does not exist.
The practical responses are all forms of telling the optimiser to be less confident:
| Approach | What it changes |
|---|---|
| Constraints | Cap any single weight, forbid shorting, set sector limits. Crude and effective |
| Shrinkage | Pull the covariance estimate toward a simpler structure, trading a little bias for much less variance |
| Black-Litterman | Start from market-implied equilibrium returns and adjust only where you hold a genuine view |
| Risk parity | Ignore expected returns entirely and equalise each holding’s risk contribution |
| Equal weighting | Assume nothing at all, and beat naive optimisation more often than is comfortable |
Equal weighting deserves its place on that list. It is embarrassing for the field how competitive a portfolio of “the same amount in each” is against sophisticated optimisation once estimation error is accounted for. If your optimiser cannot beat equal weights out of sample, it is not adding value.
What is actually available to diversify with
Before the theory is useful you need building blocks whose returns are driven by genuinely different things. The practical set is almost always smaller than the fund industry’s product count suggests, and the examples below use what is readily available to an investor in India.
Domestic equity is the core, and its internal diversification is weaker than it looks. Large, mid and small cap Indian equity are different in volatility but driven by largely the same domestic growth and liquidity cycle. Spreading across three market-cap buckets reduces single-stock risk and does much less about market risk.
Debt is the classical diversifier. Government securities and high-quality corporate debt respond to rate expectations rather than earnings, which is a different driver. The important distinction is credit quality: low-rated corporate debt tends to fall alongside equity in a crisis, which is exactly when the diversification was meant to arrive. Duration and credit are two separate decisions and are often conflated.
Gold has historically behaved differently from Indian equity and carries the additional property of rising when the rupee weakens. Its long-run real return is unremarkable, so it earns its allocation through behaviour rather than returns.
International equity is the genuine diversifier most portfolios lack, because it introduces both a different economic cycle and currency exposure. Note that the currency effect cuts both ways and often dominates the equity return over shorter horizons.
The honest summary is that a portfolio of six Indian equity funds is one bet held six ways. Real diversification requires holding things that will disappoint you at different times, which is precisely why it is uncomfortable to maintain.
Putting it into practice
Three further specifics change the calculation here.
Index concentration is higher than it looks. The NIFTY 50 is weighted by free-float market capitalisation and, as the index methodology sets out, membership is drawn from a liquid large-cap universe. The result is meaningful sector concentration, financials in particular. Holding several large-cap funds is not diversification if they all track roughly the same names. Check what you own by underlying exposure, not by number of funds.
Correlations within Indian equities are high. Diversifying across domestic stocks does less work than diversifying across asset classes or geographies. The genuine diversifiers, debt, gold, and international equity, are also the ones investors most often skip because they have been the slower performers.
Costs and taxes change the rebalancing arithmetic. Every rebalance is a taxable event and carries transaction costs, so the theoretically optimal frequency in a textbook is more often than the practically optimal frequency here. Modelling that honestly is part of quantitative analysis.
Rebalancing
Portfolios drift. Winners grow into a larger share, which quietly raises risk exactly as the winning asset becomes more expensive.
| Method | Trigger | Trade-off |
|---|---|---|
| Calendar | A fixed schedule, quarterly or annually | Simple, but trades when nothing needs trading |
| Threshold | When a weight drifts beyond a band | Responsive, at the cost of more transactions |
| Combined | Check on a schedule, act only if outside the band | Usually the sensible compromise |
The hard part of rebalancing is not mechanical. It requires selling what has done well to buy what has done badly, at the moment that feels least sensible. Automating the decision, or at least writing the rule down before the drift appears, is the only reliable way to keep doing it.
Measuring whether it worked
Return alone tells you nothing. The measures worth tracking are risk-adjusted: return per unit of volatility, and the depth and duration of the worst drawdown. A portfolio that returned 14 percent with a 40 percent drawdown is not obviously better than one that returned 11 percent with 15 percent, and the second is far more likely to be held through to the end.
Attribution matters too. Decomposing returns by asset class and by factor tells you whether the result came from the allocation you designed or from a single position that happened to work. Those look identical on a statement and are entirely different as evidence.
One caution on all of these measures: they need a meaningful sample before they mean anything. A Sharpe ratio computed over eight months of a rising market is a description of that market, not of your process. Judge an allocation across at least one full cycle, including a period you would rather forget, because that is the only stretch that tests whether the diversification you designed was real.
Where to go next
Sizing and the tail-risk problem behind all of this are covered in risk management. The estimation discipline that decides whether any of these inputs mean anything is quantitative analysis. For extending factor thinking into models that infer their own relationships, see machine learning for stock selection, and for protecting an existing equity book rather than reallocating it, hedging with NIFTY options.
Diversification is free. Precision about diversification is expensive and usually fake. Prefer the robust answer over the optimal one.