Why AI Investing Apps Keep Getting Indian Stocks Wrong

Most AI investing apps are built for clean, richly standardized Western filings, and then pointed at India as if the data were the same. It is not, and that mismatch is exactly why so many of them produce answers that sound confident and are quietly wrong.

The failure is rarely the model. It is the data underneath. Indian equity data carries a set of local realities that a globally trained pipeline never learned to expect, and treating that data as if it were clean is the fastest way to a precise-looking, incorrect number.

The tagging is not as clean as the tool assumes

Indian companies file structured financials in XBRL, the same broad standard used globally. In theory that means a machine can read every filing the same way. In practice, filers tag the same economic concept differently, use custom extensions, or place a line item under a label a global parser does not expect.

A tool that assumes one tag always means one thing will silently pick up the wrong number for some companies and the right number for others. Nothing errors out. The pipeline just ingests a slightly wrong figure and reports it with full confidence.

Clean-looking data is more dangerous than obviously messy data, because nobody thinks to check it.

This is the quiet core of the problem. A blank field gets noticed. A plausible-but-wrong value flows straight into a ratio, a chart, and eventually a screen result, and no human ever questions it.

Restatements move the past

Indian companies restate prior-period numbers more often than many investors realize. Accounting standard transitions, reclassifications, mergers, demergers, and corrections all reshape figures that were already published.

A naive tool ingests each filing once and treats it as permanent truth. But last year’s annual report may present the prior year differently from how that prior year was originally filed. If you do not track which version of a number was known at which point in time, you end up comparing a restated figure against an original one and calling the difference “growth.”

Getting this right means being deliberate about vintages: what was reported, when it was reported, and how it was later revised. Most generic tools flatten all of that into a single value per period and lose the history entirely.

Banks are not companies with a different balance sheet

This is the mistake that most cleanly separates local knowledge from imported knowledge. A bank is not an industrial company with a bigger balance sheet. It is a fundamentally different animal, and most standard-issue ratios simply do not apply.

Consider ROCE, a favorite of generic scoring engines. Return on capital employed assumes a business deploys capital into operating assets to generate operating profit. A bank’s “capital employed” is not a meaningful concept in that sense, because deposits are its raw material and loans are its product. Apply ROCE to a bank and you get a number, and the number is noise.

The metrics that actually describe a bank are different:

• Net interest margin, the spread the bank earns on its lending
• Return on assets and return on equity, sized for a leveraged balance sheet
• Gross and net non-performing assets, the credit-quality signal
• Cost-to-income, the efficiency measure
• Capital adequacy, the solvency cushion regulators watch

A tool that ranks a bank next to a cement maker on ROCE, EBITDA margin, or ROIC is comparing two things that cannot be compared. The output looks like a leaderboard. It is a category error.

The India-specific pitfalls, in one view

Many of the traps are individually small and collectively decisive. A few of the most common:

Pitfall	What a naive tool does	Why it matters
Inconsistent XBRL tagging	Reads one tag as one concept	Picks up wrong or missing line items for some filers
Frequent restatements	Treats each filing as final truth	Compares restated vs original numbers as “growth”
Banks vs non-banks	Applies ROCE, EBITDA, ROIC to banks	Meaningless ratios and false rankings
Excise duty in revenue	Takes gross revenue at face value	Overstates the top line and distorts margins
Bonuses and splits	Uses raw share count over time	Breaks EPS, book value, and per-share history
Promoter holding and pledging	Ignores it entirely	Misses a structural risk and ownership signal

Notice the pattern. None of these require a smarter model. They require someone who knew the trap was there before the data was ingested.

Excise duty, corporate actions, and ownership: three concrete traps

Take an oil marketing company as a neutral illustration. A large slice of what shows up near its reported revenue can be excise duty and other levies that pass through to the government, not value the company keeps. If a tool reads gross revenue at face value without separating that out, the top line looks inflated and every margin computed from it is distorted. This is not an exotic edge case; it is how that entire sector reports.

Now take corporate actions. When a company issues a bonus or splits its stock, the share count multiplies while the underlying value does not change. Per-share history depends entirely on adjusting for these events. A tool that carries a raw share count across a split will show earnings per share, book value per share, or price falling off a cliff on the split date, and may read that artefact as a collapse. The company did nothing of the sort.

Finally, ownership. In India, promoter holding and the share of promoter stock that has been pledged against loans are watched closely, because they carry real structural signal. Global tools built around dispersed institutional ownership often ignore promoter structures entirely. A concentrated, heavily pledged ownership base is a materially different situation from a widely held one, and a tool that never looks at it is blind to that difference.

The pattern behind all of it

Every one of these is the same underlying failure: a system built for one market’s conventions, applied to another market’s data, with the assumption that data is data. It is not. Financial data is only as good as the local knowledge baked into how it is read.

This is the specific ground Altys Labs works on: the Indian nuances, the tagging quirks, the bank-versus-non-bank distinction, the corporate actions, the pass-through revenue, the ownership structures. Not because they are glamorous, but because they are where confident tools quietly go wrong.

The uncomfortable truth is that a wrong number that looks clean is worse than a missing one. A gap invites scrutiny. A plausible figure invites trust, and trust in the wrong figure is how a slick interface leads a careful investor astray.

Practical takeaway

If you are evaluating any AI investing tool on Indian stocks, a few checks separate the careful ones from the confident ones:

• Ask how it treats banks. If it ranks a bank on ROCE or EBITDA margin next to a manufacturer, it does not understand the asset class.
• Check a company that has had a recent split or bonus. If its per-share history has an unexplained cliff, the corporate-action handling is broken.
• Look at an oil marketing company’s revenue and margins. If the top line looks inflated and margins look thin, ask whether excise pass-throughs were stripped out.
• See whether it surfaces promoter holding and pledging at all. Silence on ownership structure is a tell.
• Prefer tools that show their working: what number was reported, when, and how it changed. A single flattened figure per period is hiding the restatement problem, not solving it.

None of this is a stock view, and none of it is a recommendation. It is a reminder that in Indian equities, the hard part was never the intelligence layer. It was always the data, and the local knowledge required to read it honestly.