How Institutional Investors Build Financial Models

Institutional investors build financial models by combining sourced historical data, business-driver assumptions, linked financial statements, scenarios and continuous review. The spreadsheet remains the calculation engine, while research platforms and AI tools can reduce the manual work around it.

A practical modern stack might use a research platform to organise the evidence, Claude to help construct and audit the workbook, and Excel or another spreadsheet to preserve explicit calculations. The analyst still owns every assumption and the final sign-off.

The goal is not to replace Excel with a chatbot. It is to remove the hours spent finding, copying, cleaning and rechecking information so that more of the modelling day goes into understanding the business.

For decades, financial modelling was built around a simple bargain: the spreadsheet was extraordinarily flexible, but the analyst had to assemble almost everything around it by hand. That bargain produced rigorous models. It also produced long nights, fragile files and a great deal of work that looked analytical but was really information logistics.

AI changes those logistics. It does not change the need for an honest model.

What financial modelling traditionally involved

A finished model can look deceptively clean. Historical years sit on the left. Forecast years sit on the right. Revenue, margins, working capital, cash flow and valuation appear to move through a tidy set of formulas.

Getting to that point was rarely tidy.

First, collect the source material

The modeller began with documents, not cells.

There were annual reports, quarterly results, investor presentations, exchange filings, earnings-call transcripts and sometimes credit-rating reports. The latest period might use a different segment definition from the period before it. One filing might report in lakhs, another in crores. A note to the accounts might explain that a number appearing recurring was actually exceptional.

Before forecasting anything, the analyst had to answer basic questions:

• Which result is the latest?
• Is it consolidated or standalone?
• Was the previous period restated?
• Did the company change a segment definition?
• Is other income operational or incidental?
• Does a reported growth rate compare like with like?

This work often involved many browser tabs, a folder of PDFs and a separate sheet tracking where each number came from.

Then, spread the historical statements

“Spreading” means taking the reported profit and loss statement, balance sheet and cash-flow statement and laying them out in a consistent spreadsheet history.

The process was repetitive but unforgiving. A modeller copied each line, checked its unit, mapped it to a standard row and reconciled the total. If the company’s presentation changed, the history had to be recast or clearly bridged.

One wrong sign could distort working capital. One missed zero could change a leverage ratio by ten times. One copied total could disagree with the sum of its components without producing an obvious spreadsheet error.

The work was manual because the judgement was local. “Other current assets” could not simply be treated the same way for every company. A project contractor, a software exporter and a retailer carry very different operating balances inside similarly named accounting lines.

Next, rebuild the business behind the statements

A serious model does not forecast consolidated revenue by typing a growth percentage into one cell. It asks what creates the revenue.

For a retailer, the driver tree might include store count, selling area, sales density and same-store growth. For an IT-services company, it might include headcount, utilisation, billing rates, offshore mix and currency. For a lender, it might include loan growth, yields, funding costs, credit costs and capital.

The analyst had to find these operating KPIs across presentations and calls, then create a history that matched the financial statements closely enough to forecast.

This was the most intellectually valuable part of the build. It was also surrounded by the least valuable work: searching documents, retyping tables and trying to remember which quarter contained the useful disclosure.

Then came the three-statement mechanics

Once the historical base and driver map were ready, the spreadsheet became a connected accounting system.

The income statement projected revenue, costs, depreciation, interest and tax. The balance sheet projected working capital, fixed assets, debt and equity. The cash-flow statement translated accounting profit into cash and closed the model through the cash balance.

Every important line created another schedule:

• revenue by segment;
• operating costs and margins;
• receivable, inventory and payable days;
• capital expenditure and depreciation;
• debt, repayment and interest;
• tax and deferred tax;
• share count and earnings per share;
• valuation and sensitivities.

The model had to balance. Cash had to reconcile. Segment totals had to match consolidated totals. Historical formulas had to stop exactly where forecast formulas began.

This was where spreadsheet skill mattered. Good modellers used consistent colours, explicit assumptions, short formulas and visible checks. Weak models buried assumptions inside long formulas and used hard-coded plugs to make the balance sheet behave.

Finally, update everything again next quarter

The first model build was only the beginning.

When a new result arrived, the analyst copied the quarter, updated trailing numbers, rolled forward estimates, added new guidance, checked the call for changed language and rewrote the investment note. If a segment changed, historical comparability could break. If management withdrew guidance, the forecast framework had to change.

The spreadsheet could calculate instantly once the new inputs were in place. Getting the right inputs into the right cells was still the bottleneck.

Why spreadsheets survived every attempt to replace them

The traditional process was slow, but the spreadsheet itself was not the problem.

Spreadsheets have four qualities that remain unusually valuable.

They are explicit

A formula can be opened. A reference can be traced. An assumption can sit in a visibly separate cell. This makes the model reviewable by someone other than its author.

They are deterministic

If the inputs and formulas do not change, the output does not change. That is essential when a forecast feeds a valuation, an investment memo or an internal review.

They are flexible

No standard template can anticipate every business model. A spreadsheet lets an analyst build a subscriber schedule for one company, a store roll-out for another and a commodity-price bridge for a third.

They make scenarios cheap

Once the model is linked, a change in volume, pricing, margin or working capital can flow through profit, cash, debt and valuation. The sheet becomes a controlled environment for asking “what if?”

AI should preserve these strengths, not hide them behind an answer box.

Where the old process actually broke

The weaknesses sat around the spreadsheet.

The source material was fragmented. Historical data could be restated. KPI disclosures lived in prose. Model updates depended on a person remembering the entire chain. Reviewers spent time checking transcription instead of challenging assumptions.

The result was a poor allocation of analyst attention.

Traditional task	Why it was slow
Finding filings and disclosures	Information was spread across sources and periods
Building historical statements	Numbers had to be copied, mapped and reconciled
Tracking KPIs and guidance	Important context lived in presentations and calls
Writing formulas and schedules	Every workbook began with substantial mechanical construction
Updating after results	New data, commentary and assumptions had to be rolled together
Reviewing the model	Time went into finding broken links and hard-coded cells

None of these tasks is trivial. But not all of them require the analyst’s highest judgement.

That is the opening for a better workflow.

The new division of labour

The strongest setup does not ask one tool to do everything.

The research platform supplies context. Claude accelerates construction and review. The spreadsheet performs the calculations. The analyst owns the model.

Each layer solves a different problem.

What a research platform contributes: a better starting point

Financial modelling becomes faster when the analyst does not begin with an empty folder and a search engine.

A useful research platform brings public financials, filings, concalls, management guidance, company KPIs, valuation history and risk signals into one research context. Point-in-time data matters because a model should distinguish between what was reported then and what was restated later.

For the modeller, that changes the first questions.

Instead of spending the opening hours locating documents and reconstructing basic history, the analyst can begin by checking the business drivers, the accounting basis and what changed from the prior period. The underlying source still matters. The improvement is that the source and the structured research view sit closer together.

Altys is one example of this layer for Indian equities. Its workbooks can be used to examine point-in-time data, run calculations and build scenario exhibits before the result is exported into the modeller’s wider workflow. The company view keeps financials, valuation, forensics and projections in one research context rather than scattering them across unrelated tabs.

This is not a claim that a data platform can decide the forecast. It cannot. The advantage is a cleaner starting point and a shorter route back to the evidence.

What Claude contributes: an active workbook assistant

Claude can now create and edit spreadsheet files, and Claude for Excel can work inside a workbook to analyse formulas, trace references, make changes and help build financial models. Used carefully, this removes a second category of manual work.

Claude can help with tasks such as:

• laying out a first-pass three-statement template;
• translating a written driver map into schedules and formulas;
• explaining an inherited workbook before it is edited;
• finding hard-coded cells inside forecast ranges;
• tracing why a balance sheet does not balance;
• comparing formulas across columns for inconsistencies;
• creating scenario toggles and sensitivity tables;
• documenting assumptions and formula changes;
• updating a repeated workbook section after a new period is added.

These are meaningful gains because they occur inside the analyst’s existing calculation environment. The output is not merely prose about a model. It can be a workbook that remains open to inspection.

The limitation is equally important. Claude can write a plausible formula that reflects the wrong accounting logic. It can map a disclosure to the wrong line, preserve a flawed inherited assumption or make a tidy workbook that is conceptually weak.

So every AI-made change should be treated like work from a fast junior analyst: useful, reviewable and never self-approving.

What the spreadsheet still contributes: the source of calculation truth

The spreadsheet should remain the place where reported history, assumptions and outputs are separated clearly.

A practical model can use four visual zones:

Zone	What belongs there
Reported history	Sourced actual financials and KPIs
Assumptions	Explicit human inputs with dates and rationale
Calculations	Linked formulas and accounting schedules
Outputs	Scenarios, valuation, sensitivities and checks

Claude may help construct or audit every zone. It should not blur them.

If an assumption is generated during discussion, it must still become a visible input cell. If a historical number comes from Altys, its period and source basis should travel with it. If a formula is changed, the reviewer should be able to inspect exactly what changed and why.

The spreadsheet is where an opinion becomes an auditable calculation.

What the analyst still owns

The fastest workflow still fails if the modeller delegates the questions that define the model.

The human should own five decisions.

1. What actually drives the business?

Claude can suggest a driver tree. Altys can surface the reported KPIs. The analyst decides which drivers are economically important and which are merely available.

2. Which historical numbers are comparable?

A restated segment, acquisition, demerger or accounting-policy change may make a clean time series misleading. The analyst decides whether to bridge, recast or break the comparison.

3. Which assumptions are defensible?

Revenue growth, margins, working capital and capital expenditure are not facts about the future. They are positions under uncertainty. The analyst must be able to explain why each assumption exists and what evidence would change it.

4. Which scenarios matter?

A model can produce hundreds of sensitivities. Only a few represent genuine business risks. The analyst decides whether the critical uncertainty is demand, pricing, input costs, funding, regulation, execution or something else.

5. Is the model fit for a decision?

A balanced spreadsheet can still be a bad model. Accounting consistency does not guarantee sensible economics. Final sign-off belongs to the person or team carrying the consequence of the decision.

A practical modern financial modelling workflow

Here is what the combined process can look like from blank page to reviewed model.

Step 1: write the question before opening the sheet

Define what the model must answer.

Is the task to understand cash needs through a capacity expansion? Test whether margin guidance is achievable? Estimate the effect of currency on earnings? Compare valuation across three operating scenarios?

A model built without a question tends to become a historical database with a forecast attached.

Step 2: build the evidence pack

Collect the reported statements, segment history, KPIs, management guidance and relevant forensic checks. Confirm the period, basis and source for material inputs.

This can be done through a structured research platform such as Altys, provided the analyst can still trace material numbers back to their source.

Separate two ideas immediately:

• reported fact: what the company disclosed;
• analyst assumption: what you think happens next.

The distinction should survive all the way into the workbook.

Step 3: map the business in plain language

Before writing formulas, describe the engine.

For example:

Revenue = active capacity × utilisation × realisation

EBIT = revenue - variable cost - fixed cost

Free cash flow = operating cash flow - capital expenditure

Ask Claude to challenge the map. What driver is missing? Which assumption is double-counted? Does a cost move with volume, revenue or time? The conversation can expose weak logic before it becomes a hundred linked cells.

Step 4: let Claude draft the workbook structure

Create the historical statement layout, the driver schedules and the major accounting links. Ask for short formulas, separate assumption cells and dedicated checks.

Do not ask for “a model of the company” in one instruction. Build in reviewable blocks:

1. historical statements;
2. revenue schedule;
3. cost and margin schedule;
4. working-capital schedule;
5. fixed assets and depreciation;
6. debt and interest;
7. cash flow and balance-sheet close;
8. scenarios and valuation.

Review each block before adding the next. This keeps speed from outrunning understanding.

Step 5: inspect every historical tie

Historical totals should match the source. The balance sheet should balance. Closing cash should reconcile. Segment totals should bridge to consolidated figures or carry a written explanation for the difference.

Claude can find formula inconsistencies. The analyst must resolve the accounting reason.

Step 6: enter assumptions as a dated thesis

Every major assumption should have a short note:

• the chosen value;
• the evidence behind it;
• the date it was set;
• the condition that would make it wrong.

This turns the assumption sheet into a research record rather than a collection of coloured cells.

Step 7: run scenarios, not a single answer

The base case should not pretend uncertainty has disappeared.

Build a small set of economically distinct scenarios. Change the drivers that represent the real debate, then let the linked spreadsheet show the effects on profit, cash, financing and valuation.

Claude can create and audit the scenario mechanics. A research platform can help refresh the supporting evidence. The analyst decides whether the scenarios are plausible.

Step 8: update by exception after results

When a new quarter arrives, do not rebuild the entire research process.

Use the research platform to identify new filings, changed KPIs and revised guidance. Ask Claude to help add the period, compare actuals with the prior model and identify formulas or assumptions that need review. Then update only the parts of the thesis that the new evidence touched.

The model becomes a living comparison between what was expected and what happened.

Traditional workflow versus the combined stack

Stage	Traditional process	Research platform + spreadsheet + Claude
Source gathering	Search, download and label documents manually	Begin from organised public research context with source links
Historical spread	Copy and standardise statements by hand	Structure inputs faster, then reconcile them in the sheet
Driver discovery	Search presentations and transcripts quarter by quarter	Review KPIs and guidance together, then map the driver tree
Workbook build	Create schedules and formulas from a blank file	Let Claude draft reviewable blocks inside the workbook
Model audit	Trace formulas and hard codes manually	Use Claude to flag inconsistencies, then investigate them
Scenario analysis	Build toggles and data tables by hand	Generate the mechanics faster while keeping assumptions explicit
Quarterly update	Repeat document collection and model roll-forward	Focus on what changed and compare actuals with the prior view
Final judgement	Analyst	Analyst

The last row does not change. That is the point.

Where the time saving really comes from

The gain is not that a machine forecasts better because it types faster.

The gain comes from reducing four forms of friction:

1. Search friction: less time locating the relevant filing, KPI or prior guidance.
2. Transcription friction: less retyping and remapping of repeated structures.
3. Construction friction: faster creation of formulas, schedules and scenarios.
4. Review friction: quicker identification of inconsistent links, hard codes and broken checks.

Depending on company complexity and the state of the starting workbook, work that once consumed many hours or several days can be compressed substantially. A precise universal time claim would be misleading. A lender, conglomerate and single-product manufacturer do not require the same model.

The better measure is not hours saved. It is how much more of the total time is spent challenging the business assumptions rather than moving information between documents and cells.

Five controls that should never disappear

Faster modelling needs stronger controls, not weaker ones.

Keep source links beside material history

An extracted number is a claim until it is reconciled with the filing. Record the period and consolidation basis too.

Separate actuals from assumptions

Never let an AI-generated assumption enter a historical range or an extracted value sit inside the forecast without a visible boundary.

Review workbook changes

When Claude edits a file, inspect the changed cells and trace material formulas. Convenience is not approval.

Preserve checks

Balance-sheet checks, cash reconciliation, segment bridges and scenario integrity checks should remain visible and should fail loudly.

Seal important versions

Keep the model that existed before a result. Comparing the old forecast with the actual outcome is more useful than quietly rewriting history.

What not to automate

Do not outsource the final choice of revenue driver because one disclosure happens to be easy to extract.

Do not let Claude set a growth rate without an explicit argument you can defend.

Do not accept a balancing plug whose economic meaning is unclear.

Do not mistake a polished sensitivity table for a robust range of outcomes.

Do not update an old forecast in place and erase the evidence of what you previously believed.

Automation should remove clerical work and expose judgement. It should not hide judgement inside an automated workflow.

The future of modelling is still a spreadsheet

The spreadsheet is not disappearing. Its role is becoming more focused.

Traditionally, it was simultaneously a data-entry tool, filing archive, calculation engine, note-taking system, scenario manager and presentation surface. That was too much responsibility for one file.

In the newer workflow, a research platform handles more of the evidence and context. Claude helps the analyst build, understand and review the workbook. The spreadsheet concentrates on what it does best: explicit assumptions, deterministic calculations and transparent scenarios.

The analyst moves up the stack too. Less time is spent copying history. More time is spent deciding which history matters, which assumptions deserve confidence and what evidence would make the model change.

That is the real promise of AI-assisted financial modelling. Not a model without an analyst, but an analyst who reaches the hard questions sooner.

Public sources

• Altys: institutional research platform for Indian equities
• How to use Altys workbooks, company research and strategy tools
• Anthropic: Agents for financial services
• Anthropic: Advancing Claude for Financial Services

Related reading:

• Building a financial model from primary sources
• How to build a three-statement financial model
• Financial modelling with AI: what it does and what stays human
• Why point-in-time data matters

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.