Three consecutive analyses placing fair value above a falling price is a method failure, not three unlucky calls — audit the anchor before publishing the fourth

Claim

A fair value above spot is a buy signal. Publishing one, watching the price fall through it, cutting the estimate, and publishing another one still above spot is a loop, and the loop has a cause. Three iterations is enough to conclude the cause is internal.

Evidence — the agency's own TTD file

Analysis	Price	Conviction	Fair value called	What happened next
2026-03-25	$21.75	—	$18–28 (base $22–25)	−3%
2026-04-05	$22.05	[6.5]	$35–42 (+59% to +90%)	−37%
2026-06-11	$18.82	[5.5]	~$28 (+49%)	−27%
2026-08-07	$13.81	[2.5]	$6–10	—

Each report was individually reasonable. Each cut the prior estimate. Every one still landed above the market, and the market was right every time.

The four failure modes identified

Anchoring on the name's own historical multiple band. At its high TTD carried ~8.9x EV/revenue; 1.8x looked like a 79% de-rating and therefore like value. That band was earned by 20–25% growth. A growth-regime multiple cannot price a contraction-regime denominator. The correct anchor was the declining-peer cohort (Criteo, Ziff Davis, Teradata, Yelp at 0.6–2.2x EV/gross profit) — a set nobody assembled until the fourth analysis.
Over-weighting a founder's open-market purchase. The CEO's $148M buy appeared as a load-bearing bull signal in all three prior files. It was genuine and remains unsold — and is now −44%. Founders buy their own declining companies routinely. It is evidence about a person, not about a business.
Accepting a number the file itself flagged as inferred. The June market-share table (TTD 20%→15%) was a residual computed by subtraction, disclosed as such in its own footnote, and wrong in both direction and magnitude — the measured figure was 22%→31%. A number carrying its own caveat still hardens into a fact once it is in a table.
Treating GAAP FCF as owner earnings. $490M/yr of SBC was disclosed in every file and adjusted for in none of the fair-value work. It was the single largest markdown when finally applied.

How to apply

Before writing a fair value on a name with a prior file, read the prior fair value and what the price did to it. If the price went through it, the burden is on the new estimate to explain what was wrong with the old one — not merely to be lower.
State the anchor explicitly. "This multiple comes from X comparable set, chosen because Y." An unstated anchor defaults to the name's own history, which is the failure mode.
When a regime has broken, the name's own band is void, not merely stale. Test: was the historical multiple earned under growth the company no longer has? If yes, discard the band entirely rather than discounting it.
A drawdown is not an anchor. "Down 85% from the high" describes the past. See [[principle-down-a-lot-is-not-cheap]].
Never delete the superseded reports. Mark them superseded and keep the track record visible in the new file. The loop is only detectable if the prior calls remain legible.

What would falsify this

A fourth analysis that also lands above spot and is then vindicated by a recovery would suggest the earlier calls were early rather than wrong. Note that the current call ($6–10 against a $13.81 price) is the first in the sequence to land below the market — which is itself the test of whether the correction took.

[[principle-down-a-lot-is-not-cheap]]
[[pitfall-stale-entry-zone-suppresses-a-name]] — the same anchor problem in the opposite direction: a zone left above spot suppresses a name that is actually in range.
[[pitfall-gate-cleared-by-adjective-not-arithmetic]]

History