The method · 4 min · Updated 27 Sept 2026

How we compute intrinsic value, and where each method breaks

Two ways to end a forecast, a growth formula from Graham, and separate rails for banks, airlines and companies that lose money. Every dossier uses the same rules, written down here.

The standard route: a discounted cash flow

We project five years of free cash flow. Revenue starts from the growth the company has actually shown: the median of its last five year-over-year changes in the annual reports it filed with the SEC. The median ignores the one year of a merger or a spin-off. Without enough filings we use the five-year sales growth from market data. Year one is capped at 25 %, and the path runs in a straight line to 4 % by year five, roughly the nominal growth of the economy.

Margins are the company’s own: operating margin plus depreciation, so the model starts from cash earnings. Tax is the median effective rate of its profitable years, not last year’s, because a year with profits close to zero produces rates that mean nothing. Working capital uses the median of up to six years, never a single one-off year.

We discount at the cost of capital: a risk-free rate of 4.75 % (the US ten-year yield, reviewed every quarter), an equity risk premium on top, and a beta adjusted toward 1. When the reported beta is not plausible, we rebuild it from the business risk of the company’s sector and its own debt. The cost of debt depends on how many times operating profit covers interest.

Two ways to end the forecast

After year five a business still has a value, and it is usually most of the total. We compute it twice: with perpetual growth of 3 %, and with an exit multiple equal to the median EV/EBITDA of the company’s sector, measured on our own catalog. The dossier shows how much of the value comes from after year five, so you can see how much of it rests on the long run.

A third opinion: the Graham formula

Value = earnings per share × (8.5 + 2 × growth). It needs positive earnings, and it stops making sense when earnings fall fast, because the multiplier turns negative. In those cases the formula is left out, not forced. The value we publish is the average of the methods that apply, and a method that gives a value of zero or less never enters the average.

Where the discounted cash flow breaks, and what we use instead

Banks and insurers. Debt is their raw material, not a risk to discount. We value them on book value: the price-to-book that their return on equity justifies, (ROE − growth) / (cost of equity − growth), kept between 0.5 and 2.5 times book, with a price-to-earnings check.
Airlines. Leases are debt in another form, so we use EV/EBITDAR.
Technology with a record of growth. A capped multiple of revenue. The cap is deliberate: fast growers will often look expensive, because a tool that paid any multiple the market pays would only mirror the price.
Companies that lose money. What matters is why. If the filings show at least two profitable years out of the last eight, it is a cyclical: we capitalise the median margin of its own history, with no growth. If it loses money but sales grow, over five years and over the last twelve months, it is investing: a revenue multiple from its own industry, cut hard for the losses, and the dossier warns that the value rests on a margin it does not have yet. If it loses money and shrinks, we look at what it would fetch in liquidation: current assets minus all liabilities, Graham’s net-net. A negative result is an answer, not an error.

The verdict

The margin of safety is (value − price) / value, Graham’s definition. At 20 % or more a company is undervalued and in the buy zone; between 5 and 20 %, slightly undervalued; within 5 % either way, fairly valued; down to -20 %, slightly overvalued; below that, overvalued. If the value comes out more than 5 times the price, we do not publish it as a verdict: that is almost always the model failing, not a bargain.

What it does not do

It does not forecast the share price, and it does not know about news that is not in the filings yet. It is recomputed every night after the US close, and each dossier shows the date of the price and the version of the model it used.