Structured vs Unstructured Financial Data: Why Real Analysis Needs Both

Structured financial data is the information that already lives in neat tables, such as revenue, profit, and cash flow lines, where every value has a fixed place and type. Unstructured financial data is everything written in ordinary language around those tables, such as earnings call transcripts, notes to the accounts, and the management discussion. Real analysis needs both, and the unstructured half is the genuinely hard half, because that is where the reasons live and reasons do not fit in cells.

Here is the trap in one line. The tables tell you what happened. The text tells you why it happened. Most people build tools that read the tables perfectly and skim the text, then wonder why their picture of a company is thin.

What each half actually is

Structured data is anything that fits a grid. A number sits in a row (say, revenue) and a column (say, the quarter ended March), and its type is known in advance. Because it has a fixed shape, a computer can add it up, compare it across years, and line it up against another company without having to understand a single sentence. Financial statements, in their tabular form, are the classic example. So are share prices, filing dates, and shareholding percentages.

Unstructured data is everything that does not fit a grid. It is written for a human reader, not a spreadsheet. The bulk of any company’s real disclosure is unstructured: the management discussion and analysis, the notes to the accounts that explain what each headline number actually contains, the earnings call where analysts press management on the quarter, the auditor’s remarks, and the fine print of related-party dealings. The meaning is in the language, and the language was never designed to be machine-read.

There is also a large middle ground, sometimes called semi-structured data, where a table exists but its shape keeps shifting. Segment disclosures are a good example: they look like tables, but the segments get renamed, merged, and split from year to year, so the neatness is an illusion. Anyone who has tried to map revenue to real business lines knows this, which is exactly the problem in revenue mapping explained and segment analysis explained.

Why the structured half feels solved but is not

It is tempting to think the tables are the easy, done part. They are labelled, they add up, and standards like XBRL exist precisely to make financial statements machine-readable. So surely the structured half is handled.

Not quite. Even the neat tables carry hidden problems: a line can be tagged wrong, a value can be missing, two companies can put the same economic item on different lines, and a footnote can quietly change what a number includes. A revenue figure with no context is just a number, and the context that makes it comparable almost always lives in the unstructured text next to it. That is the argument in why XBRL is not enough: structured tags get you a shape, not an understanding.

So the honest position is that the structured half is necessary and looks tidy, but it is only half a picture, and its tidiness can lull you into trusting numbers you have not read the notes for.

Why the unstructured half is the hard half

Now the difficult part. Unstructured data resists every shortcut that makes structured data easy.

• It has no fixed place. A key admission about a slowing order book might sit in the third paragraph of the outlook section in one filing and inside a footnote in the next. There is no row to look it up in.
• The same idea is said a hundred ways. “Demand was soft,” “we saw some moderation in offtake,” and “volumes were below our internal expectations” can all mean the same thing. A person reads through the phrasing. A naive machine treats them as unrelated.
• Meaning depends on context. A number in a sentence only makes sense next to the sentence before it. “Margins improved 200 basis points” is a completely different fact if the previous line said “excluding a one-time insurance payout.”
• The important part is often what is not said. Careful readers of a concall listen for the guidance that quietly got walked back, or the segment that management stopped talking about. There is no field for an absence.

This is why searching or summarising financial text with generic tools tends to fail. The structure that a search engine or a language model expects is not there, and the discipline required to pull a fact out correctly, with the exact sentence it came from, is real work. We have written about the specific way this breaks in why earnings transcripts break search: the transcript is long, repetitive, and full of numbers that only mean something in context, which is the worst possible input for a system built to match keywords.

Why you cannot skip either one

The reason this matters is not academic. The two halves answer two different questions, and you need both to understand a business.

The structured tables answer “what happened.” Revenue rose, margins fell, receivables grew faster than sales. These are facts, and they are essential. But on their own they are mute. A margin that fell could mean three completely different things, and the number alone cannot tell you which.

The unstructured text answers “why it happened.” It is in the management discussion that you learn the margin fell because a key input cost spiked for one quarter, or because the company deliberately cut prices to defend market share, or because a competitor is now underpricing it and pricing power is eroding. Those are three very different businesses hiding behind one identical number.

Consider a simple case. A company’s structured data shows profit up nicely and cash flow flat. The tables alone might read as fine. The unstructured notes might reveal that the profit leaned on a one-time asset sale and that trade receivables ballooned, which is exactly the sort of gap between profit and cash that free cash flow versus net profit exists to catch. The number said “good quarter.” The text said “read more carefully.” Only together do they tell the truth.

The table gives you the verdict. The text gives you the reasoning behind it. Trusting the verdict without the reasoning is how good numbers hide bad businesses.

What this means for anyone building tools

If most of the useful signal sits in unstructured text, and unstructured text is the hardest thing to read reliably, then the quality of any research tool is decided in exactly the place it is easiest to cut corners.

This is the deeper reason that data quality beats model quality. A very capable model reading a thin, tables-only view of a company will produce confident, well-written, and shallow analysis, because it never saw the reasons. A more careful setup that actually reads the filings and the calls, and keeps track of which sentence each claim came from, will be slower and less glamorous and far more honest. That trade-off is the whole argument in data quality beats model quality, and it is why generic retrieval over financial documents so often disappoints, as covered in why RAG fails for equity research.

The practical stance that falls out of this is simple to state:

• Treat the tables as necessary but not sufficient. Always read the note behind the number before you trust it.
• Treat the text as the main event, not an afterthought. The reasons, the risks, and the walk-backs live there.
• Keep the source attached to every extracted fact. A number pulled out of a filing without the sentence it came from cannot be checked, and an unverifiable financial number is a liability, not an asset.

What to take away

Structured data is the neat, tabular part of a company’s disclosure, and it is essential but easy. Unstructured data is the written part, the concalls and notes and discussion, and it is where the reasons live, which makes it both the more valuable half and the far harder one to work with. Real analysis, human or machine, needs both: the table to know what happened and the text to know why. Any tool or process that reads the tables beautifully and skims the text is describing a business it has not actually read.

The habit is worth repeating: never trust a number you have not read the words behind. The distance between what the tables say and what the text explains is usually the most important thing you can learn about a company, and it is the one thing a grid alone will never show you.

This article is educational. Altys Labs is not a registered research analyst or investment adviser, and nothing here is investment advice or a recommendation to buy, sell, or hold any security.