"Down a lot" is not "cheap"

Claim

A drawdown is a statement about the last few months. A multiple is a statement about what you pay for the earnings. They are not the same measurement, and in a post-melt-up market they point in opposite directions.

The arithmetic: a name that ran +400% off a low and then fell 40% is still up ~200%. The drawdown is real and the name is expensive. "It's down 40%" describes the first fact while quietly implying the second.

Evidence

From the July 2026 tech sweeps — every field examined was simultaneously far off its high and far above its own historical band:

Field	Off high	Off low	Verdict
Semi test / ATE	−15% to −35%	+110% to +582%	11 of 11 above own band
Optical / photonics	−20% to −60%	large	10 of 13 above own band
Robotics (quality names)	material	large	32–35x earnings
Japan AI names	−8% (FX artifact)	large	Advantest 60x P/E, 28x P/B

Sharpest single case: TER traded 76% above its own FY2025 high multiple while screening as a large drawdown. LITE sat at 2.6x its own recent-high P/B.

What would falsify this

A name that is down heavily and below its own 5-year median multiple and whose earnings base has not permanently reset. That combination is a genuine value signal — this principle exists to find it, not to rule it out. The failure mode it guards against is treating the drawdown alone as sufficient.

⚠️ Companion clause — the band is only valid if the denominator's trajectory is comparable

Added 2026-07-31 from the RBLX case — the first time this principle was at risk of being misapplied bullishly.

The July 2026 sweeps used this principle to reject names. RBLX is the inverse: it passes on its face — 4.5x P/S against a 5.3–7.1x range of prior-year troughs, below the low of every year it has traded. A clean pass, and a genuine one.

It still failed, on the third condition above ("earnings base has not permanently reset"), and the mechanism generalises:

A historical multiple band earned on a growing denominator is not a valid benchmark for the same multiple on a shrinking one. Comparing them is a category error.

Two consequences worth internalising:

1. A low multiple on a falling base is not a floor — it is a race. At a constant price, RBLX's EV/Bookings travels 2.85x → 3.09x → 3.64x as bookings roll over. The stock gets more expensive by doing nothing. Cheapness that decays is not cheapness.
2. Check which way the denominator is moving before crediting a band pass. The band test answers "is this below what it has historically been worth?" It does not answer "is the thing being multiplied still the same thing?"

Practical rule: the band test is necessary but not sufficient. A pass earns further work; it does not earn a position. Pair every band pass with a trajectory check on the denominator, and prefer a forward denominator where the reported one is known to lag (RBLX's GAAP revenue amortises bookings over 27 months, so the reported figure was describing a booking level the company had already stopped achieving).

Related

Companion filter: [[principle-story-vs-revenue]]. A drawdown screen produces both errors at once — expensive names that look cheap, and story names with no revenue.

[[pattern-ranking-function-as-policy-variable]] — the RBLX case's other half, and the reason its denominator reset: a growth vintage produced by an extraction-weighted ranker was never a run-rate to build a band on.

[[pattern-ai-levered-fields-trade-above-own-band]] — the July 2026 pattern this principle generated. RBLX is the clean counterexample: genuinely below its band, and still not a buy.

History

• 2026-07-29 — established across three sweep rounds; the single highest-yield filter of the exercise. Applying it first would have eliminated most candidates before any fundamental work was done.
• 2026-07-31 — companion clause added from RBLX. First case of the principle passing and the name still failing, which exposed that the band test is necessary but not sufficient: the denominator's trajectory has to be comparable for the comparison to mean anything.