Two Ways Companies Die: Five More Pre-Registered Tests

Second pre-registered batch · N1–N10, five tested

Last week, we registered four hypotheses about what our Blitz Index predicts — and published the results of all three we could test: all rejections. This week, we registered ten more. Five of them could be tested immediately with data already in hand. Here is what happened.

The same rules apply as before: every hypothesis, its definition, and its rejection criterion were fixed in writing before any computation. Results are reported as technical statistics of past patterns, not as buy or sell signals.

The batch came from four facts the site itself already admits: the model measures the quality of the business, not the price of the stock; its most reproducible strength is failure detection; speed has no discriminating power; and the v1.1 research agenda is public. The ten hypotheses grew out of those facts. Five could be tested now with data already in hand.

The scoreboard

Three provisional survivals, two rejections. But the scoreboard is the least interesting part. What these five tests found together is a map of how companies die — and it has two distinct territories.

Death one: speed without a moat

The Speed–Moat Gap (N2) measures how far a company's normalized speed score runs ahead of its normalized moat score. The intuition came from WeWork: speed 19/25, moat 18/60 — a 46-point gap.

Across 55 blind observations, the top-quartile gap group failed 69.2% of the time (9 of 13), against 7.7% (1 of 13) for the bottom quartile. In the 20-company blind backtest panel, the separation was total: 100% versus 0%. The five widest gaps in the whole sample — Blue Apron, Groupon, WeWork, Peloton, MoviePass — all failed. When speed outruns the moat by that margin, collapse is close to certain.

But we also found the indicator's blind spot, and we report it. In the 2021 peak panel, the gap could not discriminate at all (top 60%, bottom 60%, correlation ≈ 0). The failures there — Baidu, Zillow, PayPal — had negative gaps: solid moats at insane prices. The gap catches one kind of death and is blind to the other.

The gap is a tail signal, not a linear ranker. The five widest gaps (46–51 points) all ended in failure, but below that extreme the field mixes: Netflix carried a gap of 34.7 and returned +1,240%, Amazon's gap was 11.3 and it returned +665%. The indicator reads like a fire alarm for the extreme case — it does not grade the middle.

Death two: the expensive moat

That is where Quality per Dollar (N1) comes in. The ratio of score to P/S at the test point reached a pooled Spearman correlation of +0.413 with forward three-year returns across 80 observations — stronger than the score alone (+0.234) or the price alone (+0.335).

And again, the honest qualification: the claim said "both regimes." At the 2021 peak, the ratio ranked losses correctly (ρ = +0.412); at the 2022 trough, it said nothing (ρ = −0.073). Even at the peak, the top tier still lost — median −63.6%. It is a loss-reduction ordering, not a rescue plan. Clustering by company (the same company appearing cheap once and expensive once) shrinks the pooled correlation to +0.06, the same overlapping-panel structure we disclosed in the slope test.

The pooled tertile split tells the same story with numbers: low ratio, median −47.7%; mid, +51.2%; high, +89.0%. But the regime split is the honest headline. At the 2021 peak, the in-panel tertiles went from −73.8% (low) to −63.6% (high) — correctly ordered, both deeply negative. At the 2022 trough, the order inverted: low +112.1% beat high +97.7%.

Taken together, the two results form something close to a complete picture: the gap detects "speed without moat" collapses; the price ratio detects "expensive moat" collapses. Each covers the other's blind spot. Neither is confirmed — both are provisional, pending replication on fresh samples — but as a pair they are the strongest signal our research track has produced.

The counterintuitive rejection: margin dilution is normal for winners

The Margin Elasticity test (N5) asked whether a single number — how gross margin moves as revenue doubles — could match the full 60-point moat score at classifying success and failure. It lost decisively: 0.390 accuracy versus 0.615.

The reason is worth remembering. Of 25 test windows where elasticity was negative, 24 were successes. Tesla, Shopify, Square, MercadoLibre — their gross margins diluted as they scaled, and they won anyway. Slight margin dilution during scale-up is the normal state of a successful blitzscaler (median elasticity −1.51). The hypothesis's intuition ran exactly backward in-sample: margin expansion with scale is a nice confirmation, but margin dilution is not a disproof.

The detail is instructive: 14 true positives, 2 true negatives, and 24 false negatives — the indicator called success "failure" whenever margins thinned. And the metric has a pathology at the edge: for the one bankrupt case with shrinking revenue, both numerator and denominator turned negative, producing a huge positive elasticity for a dead company. The sign convention was kept as registered; the pathology is recorded, not repaired.

This rejection sharpens our series' running thesis. Our score measures the quality of the business; it is the moat, not the P&L path, that discriminates. Winners and losers bleed margin alike on the way up. What separates them is what the money buys.

Institutions keep their own clock

The Post-Threshold Drift test (N10) shared its events with our earlier 65-crossing test (rejected on returns) but asked a different question: does the institutional shareholder base expand after a crossing — a "discovery lag"? Using our local 13F backfill, the answer is no. Holder counts did rise in absolute terms — but less than the same stocks showed in non-crossing windows (+23.5% vs +52.1% at +4 quarters). All four indicators came in weaker than the control. Crossing 65 is not the starting gun for institutional discovery.

A data limitation is disclosed: processed 13F only covered 2016Q3–2021Q2 in the workspace, so six of nine events could not be tested. The verdict rests on three verifiable events.

The three verifiable events tell the same story individually: MELI 2017 (378 holders, +23.5% at +4 quarters), NVDA 2017 (1,085 holders, +20.1%), MELI 2019 (534 holders, +61.6%) — all rising in absolute terms, all rising less than the same stocks in their non-crossing windows.

The honest weak pass

Clean C at the Trough (N7) technically survived — clean-C median +112.1% beat dirty-C +64.0% — but its persuasiveness is thin. The single largest classified winner, Spotify at +570%, was a dirty C (demerit −8), and the B-grade group's median (+119.8%) was the highest of all. What won at the trough was not "cleanliness" but the trough itself: cheap entry multiples, the same consistent story from our earlier tests. We record the criterion as met and the hypothesis as weak.

The detail matters too. The clean-C group was PRLB (+267.6%), PSTG (+151.4%), EXAS (+112.1%), ZM (+44.3%), and Z (−9.9%). Meanwhile the two largest trough winners of all, Twist (+571.1%) and Twilio (+491.2%), could not even be classified — their demerit components were never recorded. The strongest data point for the "C at the trough" intuition sits outside the test.

What remains on the board

Ten hypotheses were registered; five were testable now. The rest wait on data accumulation: the moat-delta variant (N3) needs forward re-scoring quarters, the contract-chain (N4) and backlog-conversion (N9) studies need new disclosures, the regime-dependence (N6) needs the 150+ observation panel, and the re-blitz event study (N8) needs its event list frozen before any returns are examined.

And one discipline note. Ten hypotheses tested at once means chance can produce a survivor. Every pass in this batch is marked provisional until reproduced on a separate sample — including the two we find most promising. That is not modesty; it is the rule we registered.

A word on the work itself. Every test logged its exclusions instead of hiding them: 15 observations missing score or P/S inputs in N1, 5 without component scores in N2, 19 without computable financials in N5, 3 unclassifiable C-grades in N7. The EV/S variant of the price ratio could not be run at all — no balance sheets at the test points — and it is recorded as such. The two hypotheses we find most promising are exactly the two we refuse to call confirmed. That asymmetry is deliberate: the registration rule says a pass must survive a separate sample before it earns the word.

BlitzLab scores businesses against a scholar's standard — not investment advice. No price targets, no buy/sell recommendations. Figures cited: N1–N10 result reports (2026-10-14), pre-registration HYPOTHESES_PREREG_N_2026-10-13.md. Technical statistics of past patterns; outcomes are estimates from historical data, not guarantees.