Sourcing and Stress-Testing DCF Assumptions with AI

A discounted cash flow model is only as good as its assumptions. Growth, margins, capex intensity, working capital, terminal value — every one of those inputs is a judgment call that a PM or investment committee will challenge. AI does not replace the judgment, but it does an excellent job of sourcing comparable assumptions, running sensitivities, and red-teaming your model before someone senior does it for you.

This lesson is about using AI to make your DCF defensible. It assumes you can already build a DCF — free-cash-flow build, WACC, terminal value, discounting conventions. If the mechanics of the model itself need shoring up, Financial Modeling & Valuation teaches the build; here we focus on sourcing and stress-testing the assumptions that go into it.

What You'll Learn

• How to source growth, margin, and capex assumptions from comparable companies
• How to use AI to derive a defensible WACC
• How to run multi-variable sensitivities and identify the swing factors
• How to red-team your own model with AI

A standing rule for the whole course: never trust AI to do DCF math in plain chat — it predicts numbers rather than computing them. Build the model in Excel and use AI for the assumptions and the commentary. The full set of guardrails on AI-generated numbers is in the final compliance lesson.

Sourcing Growth Assumptions

The biggest single driver of a DCF is the revenue growth path. Most analysts pull the company's own historical CAGR and pick a number. A more defensible approach: triangulate.

Prompt:

I am building a 10-year DCF for [COMPANY] in the [SECTOR] industry.
Help me triangulate a revenue growth forecast. For each source below,
give me a number with a one-sentence rationale and a confidence level
(high / medium / low):

1. The company's own 5-year historical revenue CAGR
2. Industry growth forecast from Gartner, IDC, Mordor Intelligence,
   or a major equity research house
3. Peer median NTM growth from my comp set
4. Consensus analyst growth expectation (from FactSet or Bloomberg)
5. Bottom-up build using publicly disclosed customer cohort behavior

Recommend a base, bull, and bear case growth path for years 1-5 and
years 6-10 (fade to terminal). Justify any divergence between cases.

You verify each number against the original source, but the AI did the structuring for you.

Sourcing Margin Assumptions

Margins are sticky but they bend with scale. AI helps you find precedents:

For [COMPANY], I am forecasting gross margin expansion from [X]%
today to [Y]% in year 5. Identify three publicly traded peers that
underwent a similar margin expansion in the past decade. For each,
report:

- Starting and ending gross margin
- Time period over which expansion occurred
- The disclosed drivers (scale, mix, pricing, automation, etc.)
- Whether the expansion was sustained or reverted

Conclude with whether my [X]% to [Y]% expansion looks aggressive,
in-line, or conservative versus those precedents.

This kind of precedent search used to require browsing through 40 annual reports. With AI it is a 10-minute task — though you still verify the precedents are real, since AI will sometimes invent a clean-looking comparable that does not exist (the verification routine is in the final compliance lesson).

Building a Defensible WACC

WACC is where junior analysts get tripped up in committee. There are five inputs (risk-free rate, equity risk premium, beta, cost of debt, capital structure) and a defensible answer for each. Prompt:

Help me build a WACC for [COMPANY]. For each input, give me the
typical analyst convention and the current market value:

1. Risk-free rate: which Treasury tenor to use, and current yield
2. Equity risk premium: Damodaran's current US ERP estimate
3. Beta: 2-year vs 5-year, weekly vs monthly, levered vs unlevered
4. Cost of debt: current corporate spread for the company's credit
   rating, plus the risk-free rate
5. Capital structure: book vs market, target vs current

For each, flag the most common analyst mistake. Conclude with a
WACC range (low / mid / high) and where this WACC should land
versus sector medians.

You still pull the current numbers from your data source — AI is not a real-time market data feed. But you have a structured WACC build and a sanity check.

Stress-Testing With Multi-Variable Sensitivities

A football-field sensitivity table is more useful than a one-variable sensitivity. Build it in Excel, then ask AI to interpret:

I built a sensitivity table for [COMPANY]'s DCF showing implied
share price across two variables:

- Terminal growth rate: 2% to 4%
- WACC: 8% to 11%

The grid is attached. Tell me:

1. Where is the model most sensitive? Which combination produces
   the largest swing in fair value?
2. Where are the unrealistic corners? Identify any cell where the
   implied valuation diverges materially from the company's current
   trading multiple.
3. What is the "consensus" cell — the combination of assumptions
   that produces an output closest to current share price? Reverse
   engineer what the market is implying.

Show your work as a short markdown analysis.

The "what is the market implying" question is the most useful one. It reframes the DCF from "what should the stock be worth?" to "what does today's price imply about the future?" — which is the question portfolio managers actually ask.

Red-Teaming Your Own Model

Before sending a DCF to your senior, run an AI red team:

You are a skeptical portfolio manager reviewing my DCF on [COMPANY].
The model assumes:

- 5-year revenue CAGR of [X]%
- Terminal EBITDA margin of [Y]%
- Capex as % of revenue settling at [Z]%
- Terminal growth of [G]%
- WACC of [W]%

Attack this model. List the five questions a tough PM would ask in
priority order, and for each:

a) What evidence would I need to defend the assumption?
b) What is the strongest counter-argument?
c) How much does the answer move the fair value?

End with a one-sentence verdict on whether this model would survive
investment committee.

This is the highest-leverage prompt in the entire course. It forces you to find your model's weak points before someone else does.

Terminal Value Sanity Checks

Terminal value is usually 50-80% of total enterprise value in a DCF, which means it deserves attention. Two sanity checks AI can run for you:

For my [COMPANY] DCF, the terminal value implied EV/EBITDA multiple
in year 10 is [X]x. Tell me:

1. Is that multiple realistic for a mature business in this sector?
   Compare to current trading multiples for mature comparables.
2. Does the perpetuity growth rate of [G]% exceed long-run nominal
   GDP growth for the company's primary geography?
3. What share of total EV does the terminal value represent? Is that
   ratio normal for this kind of business?

Flag any of these as a problem.

If your terminal value implies a 30x EV/EBITDA multiple on a mature consumer staples company, you have a problem AI will catch immediately.

Documenting Assumptions for the Committee

Last step: turn your assumption work into a one-pager. Prompt:

Based on our analysis, produce the assumptions one-pager for the
[COMPANY] investment memo. Sections:

- Revenue: base case path, drivers, key sensitivities
- Margins: trajectory and precedents
- Capex and working capital: as % of revenue, justification
- WACC: build and sensitivity
- Terminal value: assumption and sanity check

Each section 60-80 words. Investment-committee tone. Cite specific
numbers but do not invent them — use only numbers from our prior
conversation.

You have a defensible memo section that took an hour instead of half a day.

Key Takeaways

• Triangulate growth from at least three sources — historical, industry forecast, peer median, consensus, bottom-up
• Use AI to find margin precedents from comparable companies' historical expansion paths
• Build WACC by walking through each input with AI as your checklist
• The "what is the market implying" reframe is the most useful sensitivity question
• Always red-team your model with AI before committee — it finds the questions a PM will ask