tryaltys.ai Request access
altys.ai
Research Workflow

Building a Financial Model from Primary Sources

Build a model from the filings themselves: pull the reported statements, read the notes, rebuild the history, then drive it with segment and KPI assumptions.

To build a financial model from primary sources, you take the numbers straight from the company’s own filed statements, read the notes that explain how those numbers were built, rebuild the reported history yourself, and only then forecast it forward using assumptions tied to the real drivers of the business. The alternative, starting from a data vendor’s tidy pre-chewed spreadsheet, saves an afternoon and quietly costs you the one thing a model is supposed to give you, which is your own understanding of how the company actually makes money.

This is not modelling snobbery. A vendor sheet has already made dozens of judgement calls before you opened it: which line items were combined, how exceptional items were treated, how a demerger was stitched into the history, how segments were mapped to a standard template. Each of those calls is defensible, and each one buries a decision you should be making. When you source from the filings directly, you see every one of those decisions, and you get to make them on purpose.

Start with the filed statements, not the summary

The primary sources are the documents the company is legally on the hook for: the annual report, the quarterly results, and the notes to the accounts. Everything else, screener sites, broker sheets, database exports, is a derived product of these. Derived products are useful for speed and for cross-checking, but they are copies, and copies drift.

Pull the reported profit and loss, balance sheet and cash flow statement exactly as filed. Do not standardise them yet. At this stage you want the company’s own presentation, because the way a company chooses to group its own numbers tells you how management thinks about the business. A retailer that reports revenue net of a particular levy, or a lender that leads with net interest income, is showing you its own mental model. Note it before you flatten it.

If you want the intro to this material, our guide on how to read an annual report covers the document structure. Here we assume you can find the statements and go straight to what you do with them.

Read the notes before you trust the numbers

The three statements are the headline. The notes are where the headline is either confirmed or quietly undermined. This is the step vendor sheets skip entirely, and it is the step that separates a model you can defend from a model you merely built.

The notes tell you what a number is made of. A single “other income” line can be interest on cash, a foreign exchange gain, a one-time sale of an asset, or a fair-value swing on an investment. For a forecast, those behave completely differently: interest on cash recurs, a one-time asset sale does not. Treat them the same and your forward numbers inherit a distortion you never saw.

The notes also flag the adjustments that matter most for trust:

  • Changes in accounting policy, which can move profit between periods without anything real happening in the business.
  • Exceptional or one-off items, which you usually want to strip out to see the underlying run rate, and then keep visible so you can add them back when they are real.
  • Restatements of the prior year, where the comparable figure printed today is not the figure that was printed a year ago.

That last point deserves care. When a company restates history, for a demerger, a discontinued operation, or a redefined segment, the neat ten-year series you assemble today is not the series that was knowable on each past date. Building on “as reported today” data can flatter a backward-looking test, because the past has been tidied with knowledge that did not exist at the time. This is the heart of why point-in-time data matters, and it is one reason the primary-source discipline is not optional if you care about honesty.

Rebuild the history yourself

Now you standardise, but on your own terms. Rebuild several years of the statements into a consistent format where every line means the same thing in every year. This is tedious and it is the point. By the time you have reconciled the reported cash flow to the change in cash on the balance sheet, and tied net profit through to retained earnings, you understand the company in a way no downloaded sheet can give you.

Two checks tell you the rebuild is honest. The balance sheet must balance in every historical year, assets equal to liabilities plus equity, because that is an accounting identity, not a target. And the cash flow must reconcile: the closing cash it produces has to match the cash on the balance sheet. If either fails, you have mislabelled or misplaced a line, and the model is telling you so before you have forecast a single quarter.

A history you rebuilt yourself is a history you can trust. A history you downloaded is a history you are hoping about.

Drive it with segments and KPIs, not the headline

Here is where primary sourcing pays off, because the filings give you the pieces to model the business rather than the total. A consolidated topline is an average of very different economics, and forecasting the average is how you end up with a number you cannot explain.

Reliance Industries is the clearest teaching case, because its own segment disclosure for the quarter ended September 2025 shows how different the parts are:

SegmentRevenue (₹ crore)Segment EBIT margin
Oil to Chemicalsabout 1,60,600about 9%
Reliance Retailabout 90,000about 6%
Jio (digital services)about 42,700about 52%
Oil and Gas (upstream)about 6,100about 83%

Read the table and the modelling lesson writes itself. The largest segment by revenue, Oil to Chemicals, runs one of the thinnest margins. The fattest margins, Jio and upstream, are far smaller by revenue. A single consolidated topline blends all of it into one indistinct number, and any assumption you apply to that blend is applying one growth rate and one margin to four businesses that have almost nothing in common.

So you model each segment on its own driver. Oil to Chemicals moves with volumes and refining and petrochemical spreads. Retail moves with store count, area, and sales density. Jio moves with subscribers and average revenue per user. Upstream moves with production volume and realised price. Each of those is a KPI you can find, forecast, and later check against reality. The consolidated revenue and profit then fall out as the sum of the parts, which means your total is explainable line by line rather than asserted.

This is the same principle as revenue mapping, the discipline of tracing where revenue and profit actually sit inside a company, applied one step further into a live model. And it is why the three-statement mechanics, the plumbing that links P&L, balance sheet and cash flow, are only half the job. The plumbing is generic. The segment and KPI assumptions that feed it are where the real work and the real judgement live.

Why the discipline holds

Sourcing from filings is slower on day one, and it is the reason experienced analysts still do it. Three things come out of it that a shortcut cannot buy.

You catch the adjustments. Because you read the notes, you know which numbers are clean run rate and which are dressed up by a one-off, so your forward view is not quietly built on a figure that will not repeat.

You own the assumptions. Because you rebuilt the history and split it into segments, every forecast line has a driver you chose and can defend, rather than a growth rate inherited from someone else’s template.

You can update it fast. The first build is the slow part. Once the structure is right, each new quarter is a matter of dropping in the freshly reported segment numbers and checking them against what you expected. Building the model is the research. Maintaining it is the payoff.

The forecast itself, turning those clean drivers into a forward number with a range and a stated set of assumptions, is a separate craft. That is the subject of a companion piece on how analysts forecast revenue before earnings. But the forecast is only ever as trustworthy as the history it stands on, and the history is only trustworthy if you built it from the source. A model is not a spreadsheet you filled in. It is an argument about how a business works, and an argument is only as good as the evidence underneath it.

Frequently asked questions

What does building a model from primary sources mean?

It means constructing the model from the company's own filed statements and the notes behind them, rather than from a data vendor's pre-summarised sheet. You pull the reported numbers, understand how they were built, rebuild the history yourself, and only then forecast.

Why not just use a data provider's spreadsheet?

A vendor sheet is a convenience, but it hides the choices made to standardise it: which items were netted, how one-offs were treated, how segments were mapped. Those choices are exactly where judgement lives, and where errors hide. Primary sourcing lets you see and control them.

Where do the forecast assumptions come from?

From the drivers of the business, not the headline total. You break revenue into segments or units, attach a volume and price or a KPI to each, and forecast those. The consolidated number is then an output, not an input.

How long does it take to build a model from filings?

The first build of the history is the slow part and can take a full day or more for a complex company, because you are reading notes and reconciling numbers. Once the history is clean and the structure is right, updating each quarter is fast.