---
title: "How we value a stock: twelve models, one number"
description: "The headline on a valuation page here is the median of twelve models. Which twelve, which of them count for a company, and where the rate comes from."
date: 2026-09-21
updated: 2026-09-21
tags: [valuation, methodology, basics]
author: "Allen Yan"
image: https://sweetvaluelab.com/blog/img/2b7045208faacbbb.png
lang: en
canonical: https://sweetvaluelab.com/blog/how-we-value-a-stock
---

NVIDIA closed at $218.54 on September 18, 2026. The [valuation page](https://sweetvaluelab.com/stocks/nvda) here led with a model value of $95 a share that day, a 130% premium to the model. I wrote the code that printed it, and the sentence I would like to be precise about is this one: $95 is the median of twelve models, run on one set of filed statements at one discount rate, after the results that failed a sanity check were dropped.

There are four decisions inside that sentence. Which twelve models. Which of them count for a given company. Where the 12.4% discount rate came from. What happens when a model returns something absurd. This post is those four, in order, and everything in it is a constant or a rule you can read in the code. No model is asked for an opinion anywhere in the pipeline, and running the whole thing twice on the same filings gives the same answer.

For most of a year this explanation lived on a page called `/methodology`, written in the third person with the tone of a specification. It reads better as what it actually is: a set of choices I made, including the two I got wrong the first time. Those two are section one.

![Filed statements and one discount rate feed twelve models; the weighted ones that survive a sanity band are reduced to a median and a range](https://sweetvaluelab.com/blog/img/d54b6c73b5c92979.svg "Schematic of the reference valuation. Twelve models run on the same filed statements and the same WACC. Primary models and cross-checks carry weight; sanity checks and reference models are drawn on the page and left out of the median. A result at or below zero, or outside 0.05×–20× the price, is dropped before the median is taken. The band on the page is the range of what survived.")

## 1. The headline used to be a single model

The first version led with one model: the five-year FCFF discounted cash flow, growth exit. It is the model a bank analyst would build, so leading with it felt defensible. It failed in two directions.

When it returned zero, the page had no headline at all. Boeing, Caterpillar, Micron and Regeneron each showed nine or ten model results down the page and an empty space at the top, because the one model chosen to speak for them could not run.

When it blew up, the page led with the wreckage. A bad share count or a projected margin of 70% on a business that earns 4% goes straight through a DCF and comes out the other end as a value per share. McDonald's led with $302 million a share. Elevance led with a figure 38 times its own price. Both were on the site, in public, for anyone who opened those pages.

So the headline is now the median of the models that carry weight, and a wide sanity band drops the wreckage first. The median does the work here. An average would move toward whatever the band let through, and a headline that depends on the worst model that ran is worth nothing.

## 2. The twelve models

Every model answers a different question about the same company. They are the ones a sell-side or buy-side analyst would build in a spreadsheet, which is deliberate. There is no proprietary method here. It is the standard set, run the same way on 518 companies every night.

| Model | The question it answers | What it needs |
|---|---|---|
| Growth Exit 5Y | What is five years of free cash flow to the firm worth, if the business then grows forever at its terminal rate? | Cash flow statement, WACC |
| Growth Exit 10Y | The same, over ten years | Cash flow statement, WACC |
| EBITDA Exit 5Y | The same five years, but the business is sold at year five on an EBITDA multiple | Cash flow, EBITDA, WACC |
| EBITDA Exit 10Y | The same, over ten years | Cash flow, EBITDA, WACC |
| Revenue DCF 5Y | For a company losing money: what is it worth if its margin converges to its industry's over five years? | Revenue path, industry margin |
| Revenue DCF 10Y | The same, over ten years | Revenue path, industry margin |
| P/E | What would this company's earnings be worth at its peer group's median P/E? | EPS, peer set |
| EV/EBITDA | The same question, on EBITDA | EBITDA, net debt, peer set |
| EV/Revenue | The same question, on revenue | Revenue, net debt, peer set |
| PEG Fair Value | (earnings growth in points + dividend yield) × next year's EPS | Growth estimate, dividend |
| Earnings Power Value | What is it worth at today's normalized earnings, assuming zero growth forever? | Normalized EBIT, WACC |
| Dividend Discount | What is the dividend stream worth, discounted at the cost of equity? | Dividend history, cost of equity |

Each multiple is run on two legs, trailing and forward, and the two are averaged. The peer median comes from the company's own industry as the data provider labels it, which is a real limitation and section seven's problem.

All twelve are drawn on the page even when only five counted. The spread between them is the most useful thing on that screen. Twelve models landing within 20% of each other is a company whose value is mostly in cash flows that already exist; a three-fold spread means the value depends on which question you think is the right one, and the page cannot settle that for you.

## 3. Which of the twelve count

Before any model runs, a classifier reads the company's growth, margin, dividend yield, earnings volatility and asset intensity, and sorts it into one of eight archetypes — high growth, profitable growth, mature and stable, dividend payer, cyclical, turnaround, asset-heavy, loss-making. The page prints the one it used.

The archetype assigns each model a role. **Primary** and **cross-check** results go into the median. **Sanity checks** are single-input models kept on screen for reference. **Not applicable** means the model cannot say anything honest about this company and is not run. Some examples of how that lands:

- A loss-making company gets the two revenue DCFs as its primary models, because projecting its current expense ratios forward just perpetuates the losses. P/E is not applicable at all when EPS is at or below zero.
- A mature, slow-growing company gets EV/EBITDA promoted to primary. For a high-growth one the same multiple is a cross-check.
- A dividend payer gets the dividend discount model with weight. For everyone else it is a reference model, shown and uncounted.

The largest of these rules came out of a single afternoon of looking at insurers. A bank, an insurer or an asset manager *is* a balance sheet: its revenue is interest and premiums, its operating income has interest inside it, and EBITDA has no meaning for it. Feed that through a firm-level cash flow model and you get nonsense with a dollar sign on it — 5 to 40 times the price for every insurer and bank in the universe, Allstate and Arch and Progressive and Travelers together. So for those industries the four DCFs, EPV and EV/EBITDA are marked not applicable, and the headline falls back to earnings multiples. The proper answer for a bank is tangible book value and the return on it, which is why banks have their own [board](https://sweetvaluelab.com/tools/banks) and their own [guide](https://sweetvaluelab.com/blog/how-to-value-a-bank) instead of a DCF.

## 4. Where the discount rate comes from

Seven of the twelve models need a discount rate, and the same one is used for all of them. It is a weighted average cost of capital, built the textbook way and then deliberately flattened.

![NVIDIA's discount rate built from a 4.3% risk-free rate, a beta of 1.8 and a 4.5% equity risk premium](https://sweetvaluelab.com/blog/img/bb10676f83f497f4.svg "NVIDIA on the 2026-09-18 release. A sector median unlevered beta is re-levered at the company's own debt-to-equity, then pulled two thirds of the way from one: a raw 2.2 becomes 1.8. CAPM: 4.3% + 1.8 × 4.5% = 12.4% cost of equity. NVIDIA's $11.4B of debt against a $5.29T market cap is 0.2% of capital, so the blended WACC rounds to the same 12.4%.")

The cost of equity is CAPM: a risk-free rate plus beta times the equity risk premium. The risk-free rate is a ten-year Treasury proxy, currently 4.3%, held as a constant for the whole release so that two companies valued the same night share it. The equity risk premium is 4.5%, Damodaran's implied figure, reviewed each January.

Beta is the part worth explaining. A single stock's own regression beta is noisy, and it moves the final value more than almost anything else, so the engine builds it bottom-up: take the sector's median unlevered beta, re-lever it at this company's own debt-to-equity, then shrink it two thirds of the way toward one (0.67β + 0.33, the old Blume adjustment) with a floor at 0.3. NVIDIA's re-levered beta of about 2.2 comes out as 1.8. A quiet utility's 0.5 comes out as 0.67. The shrink makes the discount rates across 518 companies much steadier from one night to the next, and it costs you accuracy on any company whose beta really is extreme.

Three clamps sit around the projections. They are constants in the code, applied to all 518 companies alike:

- Growth rates are held between −10% and +30% a year.
- Terminal growth is set by archetype: 2.5% for a dividend payer, a cyclical or an asset-heavy business, 3.0% for a mature one, 4.0% for high growth.
- If the discount rate and the terminal growth rate come within 1.5 points of each other, the perpetuity is abandoned and the terminal value uses a 20× exit multiple, because `cash / (r − g)` with a spread of 0.2 points produces a thousand-fold multiple and no warning.

The whole WACC is then floored at 5% and capped at 25%.

## 5. The sanity band, and the page that admits it has nothing

A model result enters the median only if it is above zero and lands between 0.05× and 20× the current price. That band is enormous on purpose. It is a bug catcher, and it has no view on whether a stock is expensive: a fair value of half the price or triple the price sails through it untouched.

![A log scale of model value against price, with the 0.05×–20× band shaded and three real outliers outside it](https://sweetvaluelab.com/blog/img/0536f552ef4367a4.svg "Model value as a multiple of price, log scale. Everything from a twentieth of the price to twenty times it is kept. The three marks outside the band are real results the old headline shipped: Elevance at 38× its price, McDonald's at $302 million a share, and the zero returned for Boeing, Caterpillar, Micron and Regeneron.")

When nothing survives, the page says so. It prints "no headline model value", explains that the weighted models either returned nothing or landed outside the band, and shows every individual model result underneath anyway. A blank space where a number should be is an honest output, and it is also the thing I check first when a page looks wrong.

## 6. Why it never says buy

The gap between the model value and the price is reported as a premium or a discount to the model, and that is the whole vocabulary. The public endpoints strip the internal verdict field before the browser ever sees it. Words like undervalued are a claim about what you should do with your money, and this site does not know your tax situation, your horizon or what else you own.

## 7. What twelve models still cannot do

- **They are not twelve independent votes.** One WACC feeds seven of them, one set of filed statements feeds all twelve. When the share count is wrong, everything is wrong together, and the median hides it beautifully.
- **The peer set is an industry label.** Peer medians come from the data provider's industry classification. It puts a payment network and an insurance broker in the same neighbourhood as a bank, and when the label is wrong the multiple models inherit it.
- **The base year is one year.** A capex peak, a strike, a one-off charge or a pandemic year distorts every model that starts from it. The financials tab shows eight years so you can see what kind of year the latest one was.
- **It is a description, and the date matters.** Prices are the latest close and the statements are the latest filings. Nothing here forecasts a quarter, and a model value computed tonight has no opinion about tomorrow's open.

If the twelve models disagree wildly, the more useful exercise is usually to turn the arithmetic around and ask what the price itself assumes, which is the [reverse DCF](https://sweetvaluelab.com/blog/how-to-read-a-reverse-dcf). If the company is a bank, a REIT or a software business, the sector guides use the measures those industries are actually priced on. And if it is a company whose businesses are too different for one multiple, [Tesla's sum of the parts](https://sweetvaluelab.com/blog/tesla-sum-of-the-parts) is what that looks like done by hand.

Where the statements themselves come from, and what it takes to make eight years of filings line up into eight comparable rows, is the subject of [the next post](https://sweetvaluelab.com/blog/where-the-numbers-come-from).