AI for Portfolio Review — Exposure, Concentration and Correlation

The instinct behind diversification is sound and old: do not let one outcome decide everything. The failure is in the proxy most people use for it. Counting holdings measures how many decisions you made, not how many distinct outcomes you are exposed to. If twelve of those decisions all depend on the same underlying condition, you did not make twelve bets. You made one bet, twelve times, and paid brokerage on each.

In the Indian market this collapses in a few recognisable ways. The first is straightforward sector stacking — a bank, a housing finance company, an NBFC and a life insurer look like four different businesses on a holdings screen, and behave like one exposure to the rate and credit cycle. The second is thematic stacking, which is subtler: a cement name, a construction company, a capital goods maker and a public sector infrastructure financier are in four different sector buckets and are all fundamentally a bet on the same capital expenditure cycle continuing.

The third is factor stacking, and it is the one that surprises experienced investors. A portfolio can be spread across six genuine sectors and still be, in effect, one position in small-cap risk appetite — because every name in it is small, illiquid and owned by the same category of buyer. When that buyer steps away, sector labels stop mattering. Everything moves together for a reason that never appeared anywhere in your holdings list.

There is also a distinctly Indian structural layer that most generic portfolio writing ignores: promoter concentration. A great many listed Indian companies have a dominant promoter group, and a portfolio can accumulate several companies from the same group, or several with the same kind of governance structure, without the holder ever having framed it as a group exposure. That is a real, nameable structural fact about the portfolio, and it is exactly the kind of thing a structural review is meant to surface.

A structural review is not an opinion about your portfolio. It is an attempt to answer three specific questions, each of which has a factual answer once you have organised the data properly. Keeping them separate matters, because they fail in different ways and are fixed by different observations.

The first is exposure: what am I actually exposed to? Not which companies I own, but which underlying conditions my outcome depends on. Rate direction. Rural demand. Government capital expenditure. The rupee. Crude. Global semiconductor demand. Monsoon. A single holding usually carries several of these, and a portfolio-level view means aggregating them rather than reading them one company at a time.

The second is concentration: how much of the outcome rides on one thing? This has an arithmetic answer, and it is often startlingly different from the impression the list gives. A portfolio of twenty names where the top three are three-quarters of the capital is a three-stock portfolio with seventeen decorations. Concentration is not automatically wrong — plenty of thoughtful investors concentrate deliberately — but concentration you did not choose and cannot see is a different matter entirely.

The third is correlation: which of these move together, and why? Correlation in the strict statistical sense is measured from price history and requires actual data, which a chat model does not have. But the useful part of the question is causal rather than statistical: what shared driver would make these two positions fall on the same day? A model can reason about shared drivers from public company knowledge. It cannot compute a correlation coefficient from memory, and any number it offers for one is invented.

Keep those two ideas apart and the whole exercise stays honest. Ask the model for the structure of the relationship — the shared driver, the common input cost, the same end customer. Compute or source any number yourself. The failure mode of this entire article is a beautifully organised table in which three of the numbers were quietly hallucinated and everything downstream inherited them.

Before any of this is useful it has to be safe, and portfolio data is the single most sensitive category a retail trader is likely to hand to a third-party service. The rule that makes the rest of this article workable is simple to state: paste structure, not identity, and never paste a figure you have not verified yourself.

Structure means the shape of the portfolio — the company names and the weights, expressed as percentages. That is enough for every question in the previous section. A model can tell you that four of your holdings depend on the same credit cycle whether the position is worth ten thousand rupees or ten lakh. The absolute rupee values add nothing to the analysis and add a great deal to what you have disclosed.

What never goes in: your client or demat account number, your PAN, your broker login or any part of a credential, a contract note or holdings statement screenshot with identifiers still on it, your bank details, and — for anyone handling other people’s money — anything that identifies a client. The full treatment of this, including the difference between a consumer chat product and an enterprise-terms deployment, is in the data privacy and what never to paste article. Read it before you paste a portfolio for the first time, not after.

One more habit worth building early: strip before you paste, not while you paste. Export the holdings, open the file, delete the account-identifying columns, convert values to weights, and only then move the text across. Editing sensitive data inside the chat box is how identifiers survive — you meant to remove the account number and you removed it from the second line only.

And treat the conversation itself as a record that may persist. A portfolio review is not a one-off document; it is a description of your financial position with your name attached to the account it came from. If your setup retains chats, or trains on them, that is a standing disclosure rather than a single moment of one.

A single "review my portfolio" prompt produces a single vague answer. The structure below works better because it separates the model’s two jobs — classifying, then interrogating — and puts your arithmetic check between them, before any conclusion has been built on top of an unverified number.

The order matters more than the wording. Classification first, because everything downstream depends on it and it is the step where an error is easiest to catch — you know your own holdings, so a wrong sector or a mistaken business description is obvious on sight. Arithmetic second, done by you, because that is the step where a confident model is most dangerous. Only then the structural questions, which are where the actual value of the exercise lives.

Expect to correct the classification. A model’s picture of an Indian mid-cap may be several years old, and companies change: a business that was two-thirds one segment when the model last read about it may be something quite different now. Treat every classification as a claim to check against the company’s own latest reporting, exactly as you would treat any other factual claim from a model.

Finally, run the review on a schedule rather than on a feeling. Structure drifts silently — a position that ran up is now a larger share of the portfolio than you ever decided it should be, and nobody made that decision. Quarterly is enough for most holders. The point of a cadence is that it catches drift you would not have gone looking for.

Step 4 above is the load-bearing step, and it is the one most likely to be skipped because it is the least interesting. Here is why it cannot be delegated: a language model turns your numbers into text and continues that text. Short sums frequently come out right. Longer chains drift, and when they drift, the wrong total arrives with precisely the same fluency as a right one. There is no error signal in the output.

Consider the arithmetic itself, which is deliberately trivial. Suppose a holdings list has weights of 14, 11, 9, 8, 8, 7, 6, 6, 5, 5, 4, 4, 3, 3, 3, 2, and 2 percent. Those are illustrative figures invented for this example and represent nothing real. Add the top three and you get 34 percent. Add the top five and you get 50 percent — half the portfolio in five of seventeen names. Nothing in the original list announces that. It is one addition, and it changes how the list reads completely.

Now group the same illustrative weights by driver rather than by ticker. If the 14, the 9, the 8 and the 6 all turn out to depend on the credit cycle, that is 37 percent on one condition, spread across four names in what looked like four different businesses. Again: invented numbers, trivial addition, and a conclusion that no amount of staring at the holdings screen would have produced. The value of the exercise is in the grouping and the addition, both of which are yours.

If you want machine help with the arithmetic itself, use a spreadsheet or a tool that genuinely executes code over your file, and know which of the two you have. A chat model with a real code sandbox attached performs actual arithmetic; a chat model without one produces text that resembles arithmetic. The distinction is invisible in the answer and total in its consequences. The hallucination and audit trail article covers how to check which one you are dealing with.

Some concentration patterns are common enough in Indian retail portfolios that they are worth checking for by name rather than hoping the review surfaces them. None of these is a judgement about whether the exposure is good. They are structural facts that a holder should know they have chosen.

Index composition is the first. The large Indian benchmark indices are not evenly spread across the economy — financials have long been the heaviest block in the headline large-cap index, with information technology, energy and consumer names making up much of the rest. Anyone holding an index fund plus a handful of direct large-cap financials owns that block twice, and the direct holdings often feel like the diversification when they are the doubling. Check the current published index constituent weights from the index provider rather than trusting any recalled figure, including one from a model.

Promoter and group exposure is the second. Companies within the same promoter group can share funding relationships, governance practices, group-level sentiment and, at times, the same news cycle. Holding four companies across four sectors that happen to share a promoter group is a group exposure, and it will not appear in any sector-based view. Group affiliation is public information and worth an explicit column.

Market-cap band is the third, and it is where "I hold twenty stocks" most often turns out to mean "I hold one liquidity condition". Small-cap and micro-cap names in India are held by a similar buyer base and can move together with striking uniformity when that base withdraws, regardless of sector. Add a size-band column, add the weights up, and the picture is usually clearer than expected.

The fourth is the currency and commodity layer that never appears on a holdings screen at all. An exporter of services and an importer of crude derivatives sit in different sectors and are both, in part, positions on the rupee. Input-cost exposure works the same way: several unrelated-looking manufacturers can share one commodity in their cost base. These are causal links, which is exactly the sort of relationship a model can help you enumerate — while the sizing of them stays yours.

This is the boundary that makes everything above usable, and it is worth stating flatly. A structural review describes what you hold. It does not tell you what to hold. The moment the output becomes "reduce financials to fifteen percent" or "this position should be trimmed", you have crossed from description into personalised investment advice — and a general-purpose language model that knows nothing about your income, liabilities, time horizon, tax position, other assets or risk tolerance is in no position to be giving it.

Keep the model on the description side by instructing it explicitly. Ask what the structure is; do not ask what the structure should be. Ask which holdings share a driver; do not ask which one to sell. Ask what conditions would move most of the portfolio together; do not ask whether those conditions are likely, because that is a forecast and the model has no basis for one.

This is not merely a legal nicety. Structural information genuinely changes decisions on its own. Knowing that half of your capital rides on one condition is often all you needed; you did not require a model to also tell you what to do about it, and its instruction would have been worth less than your own judgement about your own circumstances. The valuable output of this whole exercise is a clear picture, not a set of orders.

If a decision does need outside input, that is a conversation with someone regulated to have it with you personally, who can see your full financial position and is accountable for what they say. Educational content — including this article — and a general-purpose model are both structurally unable to fill that role, whatever tone either one adopts.

The review is worth far more repeated than performed once. Portfolios drift by doing nothing at all: winners grow into a larger share than you ever chose, losers shrink out of view while still being held, and new positions get added against an old mental picture of the shape. Nobody decides any of this. It simply accumulates.

Keep the artefact, not just the conclusion. A dated table of weights by holding, by sector and by driver, stored somewhere you own, becomes a record of how the structure changed. Comparing this quarter’s table with last quarter’s tells you something no single review can: which way the portfolio is drifting, and whether it is drifting toward or away from the shape you thought you had.

Keep the prompts too. The value of a repeatable structural review is that the question stays constant while the portfolio changes — if you rewrite the prompt each time, differences between reviews may just be differences in what you asked. This is the same discipline the AI research workflow article applies to research generally, and the trading journal article applies to your own decisions.

A last note on scope. Everything here works because structural review is a reading and organising task, which is what language models are actually built to do. It stays reliable precisely because it never asks the model to know a price, forecast a move or assess a company’s prospects. Push it past that line and it degrades quickly into confident text about things it cannot see — which is the failure mode this entire module exists to help you recognise.