The Hardest Part of AI in Finance Is Not the Model. It Is the Data.

Everyone in financial AI is chasing the newest model. The harder, less glamorous truth is that the model is increasingly a commodity, and the real difficulty, along with the real edge, sits in the data underneath it.

Two teams can run the same architecture on the same universe of Indian stocks and get wildly different results. The difference is almost never the model. It is whether the numbers going in are clean, correctly scaled, correctly dated, and comparable across companies.

The Model Has Become the Easy Part

A decade ago, building a capable model was the moat. That is no longer true. Strong open and hosted models are widely available, the training recipes are published, and the gap between a good model and a great one keeps shrinking for most practical tasks.

What has not commoditized is the data. Financial statements are written by thousands of different filers, under evolving standards, in inconsistent formats, and revised after the fact. That mess does not go away when you upgrade your model. It gets amplified. A more capable model applied to dirty inputs simply produces more confident wrong answers, faster.

The uncomfortable implication: if your results are unreliable, the fix is usually not a bigger model. It is going back to the data.

The Data Problems That Actually Decide Outcomes

The failure modes in financial data are boring, specific, and lethal. Here are the ones that most often decide whether a system works.

Units and scale. One filer reports revenue in crores, another in millions, another in absolute rupees. A single mislabeled unit turns a 500 crore number into a 5,00,00,00,000 number, and any ratio built on top inherits the error. Multiply this across a portfolio and the model is learning from noise it cannot see.

As-filed versus restated. Companies restate. A number reported in one annual report can be quietly corrected in the next. If you train or test on the latest available version of history, you are using figures that did not exist on the dates you are pretending to analyze. As-filed and restated are two different datasets, and confusing them corrupts everything downstream.

Point-in-time correctness. This is the one that ruins backtests. If your dataset lets the model peek at information the market did not have on a given date, results look brilliant in testing and collapse in production. Earnings are announced with a lag. Estimates get revised. A point-in-time discipline means every input is stamped with when it actually became knowable, and nothing from the future leaks backward.

If you cannot say exactly what was knowable on a given date, you do not have a dataset. You have a leak.

Inconsistent tagging across filers. Even structured filings like XBRL are tagged by humans and systems that disagree. The same economic concept, say finance costs or other income, can land under different tags, or a filer can use a slightly wrong tag entirely. Line up two companies naively and you are often comparing things that are not the same thing.

Corporate actions. Splits, bonuses, and rights issues rewrite per-share history. A stock that “fell 50%” may simply have split two-for-one. Earnings per share, price series, and share counts all need consistent adjustment, anchored to the right dates, or your time series breaks precisely where the corporate action happened.

Sector differences. A bank’s financial statements are not a manufacturer’s. Banks have no revenue line in the manufacturing sense; their economics run on net interest income, provisions, and asset quality. An insurer runs on float and claims. Applying one generic template across a whole market guarantees that the model misreads entire industries.

Here is the same idea in one view.

Data problem	What goes wrong	Why the model cannot save you
Units and scale	Off-by-factor numbers, broken ratios	The error looks like a valid input
As-filed vs restated	History that never existed as used	Both versions are internally consistent
Point-in-time	Lookahead bias, fake backtests	The leak is invisible at inference time
Inconsistent tagging	Comparing unlike line items	Two labels, same word, different meaning
Corporate actions	Broken per-share time series	The break looks like real volatility
Sector differences	Banks read like factories	One template cannot fit all economics

Why the Model-First Teams Get Burned

The pattern repeats. A team adopts the latest model, wires it to whatever data is convenient, sees a promising demo, and ships. Then the results wobble. One company’s margins look impossible. A backtest that printed money refuses to work live. A sector comes out systematically wrong.

The instinct is to blame the model and reach for a newer one. But the model was rarely the problem. The problem was that the inputs were mislabeled, misdated, or incomparable, and no amount of model quality repairs an input that was wrong before the model ever saw it.

Meanwhile, the teams that win are often unglamorous. They spend their time on reconciliation, on catching a filer who tagged a number wrong, on making sure a split is reflected everywhere it should be, on guaranteeing that nothing from next quarter has leaked into last quarter. This work never demos well. It quietly determines whether anything built on top can be trusted.

There is a simple heuristic here. Glamorous work tends to be optional; boring work tends to be load-bearing. In financial AI, the data plumbing is the load-bearing part.

Indian Equities Make the Point Concrete

Indian filings are a good stress test for all of this. The universe is large and diverse, spanning banks, non-bank lenders, insurers, commodity processors, IT services, and consumer companies whose statements look nothing alike. Standards have evolved, and the depth and consistency of machine-readable history varies by company and by year.

Corporate actions are frequent. Restatements happen. Tagging quality across thousands of filers is uneven. None of this is exotic; it is just the ordinary texture of a real market. And every one of these frictions is a place where a careless pipeline silently produces a wrong number that a model will then treat as gospel.

The lesson generalizes beyond India, but India makes it vivid: the market does not owe you a clean dataset. You have to build one.

A Data-First Philosophy

This is the philosophy behind how Altys approaches Indian equities. Not a race to the newest model, but a stubborn insistence that the data be right first: correctly scaled, kept as-filed where it matters, stamped with what was actually knowable on each date, tagged consistently across very different kinds of companies, and adjusted properly for corporate actions.

We will not pretend this is the fun part. It is the part that decides everything else. A modest model on trustworthy data beats a frontier model on messy data, every time, in ways that only show up when real money and real decisions are on the line.

The model layer will keep improving on its own, driven by the whole industry. The data layer only improves if someone does the unglamorous work. That asymmetry is exactly why the data is the edge.

The Practical Takeaway

If you are building or buying financial AI, judge it by its data discipline, not its model.

• Ask how units and scale are validated, not just parsed.
• Ask whether the system distinguishes as-filed from restated numbers, and when it uses each.
• Ask how point-in-time correctness is enforced, and demand to see whether backtests are free of lookahead.
• Ask how it handles inconsistent tagging across filers, and how it knows two line items are truly comparable.
• Ask how splits, bonuses, and rights issues are reflected across price, share count, and per-share history.
• Ask whether banks, insurers, and manufacturers are treated with their own economics, or forced through one template.

If the answers are vague, the model on top does not matter. If the answers are precise and boring, you are probably talking to a team that will still be right when the demo is over.

The model is the part everyone can buy. The data is the part you have to earn.