A Rising Score Did Not Predict a Rising Stock

The hypothesis was seductive: a business whose Blitz Index score is climbing is accelerating — so its stock should follow. We tested it across 45 observations, pre-registered, and the correlation came out negative. The failures were not random. They point at what actually moves prices.

5 min read · 45 observations · 12 companies · Not investment advice

Why the hypothesis looked promising

It started with NVIDIA. Our 2016 blind score for the company was 66/B; the 2026 measured score is 87/A — a gain of 21 points over a decade in which the stock multiplied many times over. The intuitive leap writes itself: if a rising score captures a business accelerating, then score momentum should be an early warning for surging stocks.

The intuition behind it is worth spelling out, because it is the intuition this study is retiring. Blitzscaling is, at its core, a story about acceleration: a company that grows faster than its constraints, seizing scarce resources before rivals, should convert that momentum into compounding value. If the Blitz Index score tracks how well a company is executing the blitz, then the change in the score should track the acceleration itself — and an investor who buys acceleration early buys the surge. That chain of reasoning is exactly what we tested, and exactly what failed.

So we formalized the leap into a falsifiable claim and registered it before touching the data — hypothesis H3 in HYPOTHESES_PREREG_2026-10-13.md, the same registration that produced the profitability-inflection rejection published alongside this one. The claim: the two-year change in a company's score (Δ) predicts its forward returns more strongly than the score's level does, and Δ ≥ +10 in particular should be strong. The falsification rule was mechanical: if the Spearman correlation ρ(Δ, forward 3-year return) is at or below 0.1, reject.

The method, in brief

Forty-five observations across 12 companies, drawn from the annual estimated-score panel. For each company-year, Δ = score(year) − score(year−2). Entry was set at the first trading day after that year's 10-K (or 40-F, for Shopify through 2023) filing — pulled from EDGAR submissions, originals only, so there is no look-ahead. Returns are split-and-dividend-adjusted closes over the following three years. Windows that ran past October 2026 were excluded rather than extrapolated.

One methodological note matters for interpreting the result: the annual estimated scores were constructed with knowledge of outcomes — a hindsight bias that favors the hypothesis, since a known-good year invites a generous Δ. If a claim fails even with the wind at its back, the failure is more robust, not less.

The observation window itself deserves a note. Entries cluster in 2018–2023, which means the sample spans the COVID collapse and recovery, the zero-rate melt-up, and the 2022 bear market. It is a regime-heavy period: the single largest force in the panel is the market cycle, not any company-specific signal. That concentration is a caveat, not a flaw — but it means a larger, multi-cycle panel (the Track 2 effort) is the honest next step for anyone who wants to rescue the hypothesis.

The verdict is REJECT: ρ = −0.091 sits below the pre-registered 0.1 cutoff. Not only did the slope fail to predict — it pointed, faintly, in the wrong direction. The score's level did no better at +0.101.

The paradox: the flat and falling scores won

The observations with the smallest score changes beat the ones with the largest — by a median of +114.3% against +30.6% over three years. That is not noise around a null; it is an inversion. The names tell the story: Palantir 2022 (Δ −2, +1,517.7%), DoorDash 2022 (−3, +222.9%), CrowdStrike 2022 (+2, +264.5%), NVIDIA 2022 (−2, +654.7%).

What they share is not score momentum but timing: all were entered at the 2022–23 bear-market trough. A declining score in 2022 often meant a great business priced at despair. The slope did not do the work — the entry multiple did.

The sub-claim we most wanted to test — that Δ ≥ +10 is especially powerful — cannot be tested at all with this data. Only two observations clear that bar, and both are NVIDIA (2020: +12 → +495.8%; 2021: +19 → +341.0%). Extrapolating a rule from a single company's two best years is exactly the kind of move the pre-registration was designed to prevent. And a final sensitivity note, reported because the registration demands it: removing NVIDIA entirely makes ρ more negative (−0.216). The poster case for the hypothesis was actually cushioning its failure.

What really moved: the cycle, not the slope

This result rhymes with everything the research program has found so far. The ARK backtests showed that entry multiples dominate subsequent returns — the same 2022–23 trough effect. The profitability-inflection test found that by the time a milestone is announced, the expectation is in the price. Now the slope test adds the third verse: a business accelerating at a high multiple and the same business accelerating at a low one are not the same investment at all.

The most instructive single observation is NVIDIA 2022 itself. This is the company on which the hypothesis was founded — its decade-long score climb from 66 to 87 is our flagship example of a rising score accompanying a rising business. Yet the NVIDIA observation that did the most work in this panel is the 2022 one: Δ negative (−2), forward return +654.7%, entered at the trough. The poster case for slope-based entry succeeded through the multiple, not the slope. When your best example works by a different mechanism than the one you hypothesized, the mechanism is wrong.

One sentence for the whole study: the slope tells you how fast the business is accelerating; whether the stock explodes is decided by what price you paid for that acceleration. In this panel, the second factor was the entire story.

The cumulative picture: three rejections, one position

H2 (profitability-inflection entry) was rejected. H3 (score-momentum entry) is now rejected. H1 (threshold-crossing entry) was rejected too — crossings above the 65-point line returned a −5.9% median over one year against +35.9% unconditionally (n=7). Three honest negatives, published with their failures included, and together they sharpen what the Blitz Index claims to be. It is an instrument for measuring business quality — and the ARK backtests showed it is genuinely good at separating eventual winners from blowups. It is not, and we no longer have any evidence to suggest it is, an instrument for timing entries. A great company bought at the wrong multiple is a familiar failure mode; these studies just put numbers on two versions of it.

What remains on the registered agenda: H4 (revalidating the valuation-zone finding on a larger observation panel, the Track 2 effort), plus retests of all three rejected hypotheses once the wider 150-observation panel exists. The slope story may yet change with a wider Δ distribution; today's Δ range was narrow (−9 to +19, median +1), and a 150-observation panel could find structure this one cannot see. Until then, the honest report is the one above: a rising score did not predict a rising stock.

BlitzLab scores businesses against a scholar's standard — not investment advice. This article reports descriptive statistics of past patterns under a pre-registered methodology (H3, registered 2026-10-13, verdict: rejected). Past patterns are not predictions. No price targets, no buy/sell recommendations. Figures: GAAP, SEC-filed; estimates labeled as estimates.