INVESTSIGHT CAPITAL is a fintech and capital markets firm in Bengaluru, India. This article is part of our research library — for the official site see investsightcapital.com, or ask us a question.
The useful question about AI in investing is not whether it works. It is where the boundary sits — which decisions genuinely improve when a model is involved, and which quietly get worse.
This is the companion to our piece on how AI is transforming investment research. That one describes the gains. This one describes the edges — the places where more compute does not buy more insight, and where pretending otherwise costs money.
Where models are genuinely strong
Machines are better than people at breadth. Reading every filing in a sector, tracking hundreds of positions for drift, flagging when a portfolio’s factor exposure has changed without anyone deciding it should — this is work humans do slowly, inconsistently, and with fatigue. Models do it continuously.
They are also unmatched at surfacing the unglamorous: reconciliation gaps, disclosure changes, correlations that quietly moved. Most of the value of AI in research is not brilliance. It is coverage.
It helps to be specific about what breadth means for an Indian portfolio. A few thousand listed companies, quarterly results that land in a compressed window, exchange disclosures at all hours, and a steady stream of regulatory and macro news. No human team reads all of it. A model can, and can rank what it read by how far it departs from the expected. That ranking is the product. Everything after it is judgement.
Can AI predict the stock market?
No — not in the sense the question usually means. No model, ours included, can reliably forecast where an index or a stock will be next month, and anyone who claims otherwise is selling something. What machine learning can do is narrower: it can estimate probabilities over many small, repeated situations where the past is a fair guide to the future, and it can do so more consistently than a tired human. Even there, the edge is thin and erodes as others find it.
The reason prediction fails at the level people care about is structural, not technical. Markets are not weather. They are made of participants who read the same data, react to each other, and change their behaviour when a pattern becomes known. A model that finds a signal is, in part, the reason the signal disappears. Treat any forecast — human or machine — as one input to a decision about risk, never as the decision itself.
Where models are weak
Regime change. Statistical models assume the process generating the data is roughly stable — that the relationships in the training set will hold in the test set. Markets violate this constantly. Rate cycles turn, regulation changes, a sector’s economics shift. A model trained on the last decade does not know the decade was unusual. This non-stationarity is the single most important reason quantitative results decay.
Reflexivity. Weather does not change because a forecast was published. Markets do. Prices move on expectations, expectations move on prices, and a widely used signal changes the behaviour it was measuring. A model cannot see itself in the data.
Thin data for rare events. The events that matter most for a portfolio — crashes, liquidity freezes, policy shocks — are exactly the ones with the fewest examples. A handful of drawdowns is not a distribution. Any model’s estimate of tail risk is, at best, a well-dressed guess, and the honest response is to size positions as though the estimate is wrong.
Genuine novelty. A model learns from history. When conditions have no real precedent, its confidence is unchanged while its reliability is not — and that combination is dangerous precisely when it matters most.
Causation. Systems are exceptional at correlation and indifferent to whether a relationship is structural or coincidental. Acting on a spurious pattern is easier than ever.
Judgement about consequences. A model can estimate a probability. It cannot weigh what a particular loss would mean for a particular family at a particular moment in their life.
What can a model not judge?
Two categories of judgement stay human because the information they need does not live in a dataset.
The first is management quality and incentives. A model can score the tone of a concall and flag a change in related-party disclosures. It cannot tell you whether a promoter’s interests are aligned with minority shareholders, whether a capital-allocation record reflects skill or a favourable cycle, or whether a CFO’s departure after eleven months means what it appears to mean. Those judgements come from reading between lines, from pattern-matching across careers rather than across filings, and from conversations that leave no structured trace. In a market where promoter-led companies are the norm, this is not a minor gap.
The second is the client. A portfolio is not an abstract optimisation problem; it belongs to a person with a specific timeline, specific obligations, and a specific tolerance for watching it fall. A model can compute the allocation that maximises expected return for a given variance. It cannot know that a 30% drawdown would cause this particular investor to sell everything at the bottom — or that a wedding, a parent’s illness, or a business that needs capital changes what risk means in a given year. Temperament and circumstance are inputs no dataset contains.
Can a model be a fiduciary?
No, and this is not a philosophical point. A fiduciary is someone who is answerable — who can be asked why a decision was made, who bears the consequence of a bad one, and who can be held to a duty of care. A model has no duty, no stake in the outcome, and no capacity to explain itself in a way a regulator or a client would accept. Responsibility does not transfer to software. It stays with the people who deployed it.
Indian regulation is moving to say this explicitly. SEBI’s June 2025 consultation paper on guidelines for responsible usage of AI/ML in Indian securities markets proposes a framework in which the regulated entity remains accountable for the AI and ML tools it uses, with expectations around model governance, testing, disclosure and data protection (SEBI, June 2025). The exact rules will evolve. The principle — that the adviser, not the algorithm, answers for the advice — will not.
For an investor choosing an adviser, this suggests a simple test. Ask who is accountable when the model is wrong. If the answer is a person with a name, the firm has understood the problem.
The boundary is the strategy
The failure mode is not using AI. It is failing to say clearly which decisions it is allowed to make. Firms that get this right treat models as instruments that expand what a human can see, and keep the judgement — sizing, risk, whether to act at all — with people who are accountable for the outcome.
In our own research tooling, that boundary is written down before a model is built, not discovered after it fails. Screening, ingestion, monitoring and anomaly detection run without a human in the loop because being wrong there is cheap: a false flag costs an hour. Sizing, entry, exit and the decision not to act sit with a person because being wrong there is expensive and, more importantly, because someone has to be able to explain it afterwards.
That division of labour is deliberate in how we build. Automate the search. Preserve the judgement.
