‘Good to Great’: a zero-skill null reproduces the leap

The short answer. Jim Collins' Good to Great (2001, ~4M copies) found 11 companies that leapt from average to a 15-year run beating the market threefold, then distilled the shared traits — Level 5 leadership, the Hedgehog Concept, the Flywheel — as the discoverable causes of sustained greatness. We built the smallest model with skill switched off — firms that are all equally (un)skilled — and it reproduces the whole pattern. The "leap" is what selection on past performance plus regression to the mean looks like when there is no skill at all.

The claim. That a specific set of management traits caused a durable good-to-great transition, recoverable by studying the winners.

The catch. The 11 firms were chosen because they had already made the leap — selection on the outcome. Study only winners, with no failing-firm control, and any trait they happen to share looks causal. And firms picked for an extreme run are exactly the ones regression to the mean drags back. So the question isn't "do they share traits?" — it's whether the evidence can tell skill from luck. We can test that directly.

We measured it

The smallest null: simulate 1,400 firms as random walks with identical drift and volatility — no firm is more skilled than any other. Then apply Collins' own selection rule (mediocre for 15 years, then ≥N× the market over the next 15), and measure the next 15 years. If the "leap" needs skill, a zero-skill population shouldn't produce it.

Sanity check: across all firms with no selection, forward excess return is 0.00000.

So a population with zero skill differences produces a good-to-great cohort of the right size (9 firms at the 5× cut, vs Collins' 11), each spectacular in the selection window by construction — and then their next 15 years revert to the market. Going forward they beat the market 45% of the time: a coin flip (47% at the 3× cut). The "sustained" half of "sustained greatness" simply isn't there once you stop selecting on the past.

Why the leap evaporates

Two artifacts, one cohort. Selection on the dependent variable: pick winners and any shared trait is retrofittable, because you never looked at the firms with the same traits that didn't win. Regression to the mean: an extreme run is partly skill (if any) and partly luck, and luck doesn't repeat — so the more extreme the selection, the harder the reversion. This is the same shape we keep finding in the Crucible: a celebrated number that is a property of the measurement (here, selecting on the outcome), not of the world. The real test of "sustained greatness" is forward performance on firms chosen before the run — and that is exactly what the headline never reports.

This critique is well-established, and the honest framing is a runnable re-demonstration, not a discovery. The selection-on-winners / halo mechanism is Phil Rosenzweig's The Halo Effect (2007), which names Collins directly, and Jerker Denrell's undersampling-of-failure result (Organization Science, 2003). That a pure random walk with zero a-priori skill generates persistent inter-firm gaps and long win streaks is Denrell's Random Walks and Sustained Competitive Advantage (Management Science, 2004). And a formal statistical null test of the Good to Great / Built to Last streaks against chance — in which most of the named firms fail to clear significance — is Henderson, Raynor & Ahmed, Are "Great" Companies Just Lucky? (Strategic Management Journal, 2012). Our only addition is a small, transparent, runnable version that reproduces Collins' specific cohort and its forward collapse in one model with no data and no tuning — a receipt, not a finding.

One honest subtlety. A forward excess of ~0 does not prove the firms had no skill. In a competitive market, genuine managerial skill tends to be competed and capitalized away, so real skill also predicts a forward excess near zero (Berk & Green, 2004). That is precisely the point: winners-only selection with no forward test is non-identifiable — it cannot tell zero skill from real-but-competed-away skill. The defensible verdict is "the evidence can't decide," not "it was luck." What the null kills is the causal claim the book's design was never able to support, not the possibility that skill exists.

What this does and does not say. It does not claim management skill is zero, or that nothing in Good to Great is useful. It shows the evidence — winners-only, traits retrofitted, no forward test — cannot separate skill from luck plus selection. Tellingly, reality agrees with the null: several of Collins' "great" companies later stumbled badly — Circuit City went bankrupt (2009), Fannie Mae was taken into federal conservatorship (2008), Wells Fargo ran the fake-account scandal (~3.5M accounts by the 2017 tally; $185M in combined regulator penalties, 2016) — and a portfolio of the 11 bought at publication underperformed the S&P 500 (Levitt, 2008). That forward record is what regression to the mean predicts and what enduring skill would not. It is also an old pattern: Peters & Waterman's In Search of Excellence (1982) exemplars were widely in distress within two years (BusinessWeek, "Who's Excellent Now?", 1984), and roughly half of Collins & Porras' Built to Last "visionaries" had slipped a decade on — Collins is, in effect, his own ignored control group.

The falsifierGive the selected cohort a forward edge the null can't make: if firms chosen on a past 15-year leap kept beating the market significantly over the following 15 years (well above the ~45% coin-flip level and a positive excess whose CI excludes zero), the leap would carry real persistence and this verdict would be wrong. Better still, a pre-registered list of "great" firms, scored only on returns after the list is fixed, that beats a matched control — that would be skill the artifact can't fake.

FAQ

Does this prove Good to Great is wrong? No. It proves the study design can't support its causal claim: a zero-skill null reproduces the same 11-firm leap and shared-trait story, so the evidence can't distinguish skill from luck plus selection. The traits might still help — the book just doesn't show it.

What is selection on the dependent variable? Choosing your cases by their outcome (here, firms that already became great) and then looking for shared causes. Without the firms that had the same traits but didn't succeed, any shared trait looks causal when it may be chance.

What is regression to the mean here? An extreme 15-year run is part skill, part luck. Luck doesn't repeat, so an extreme cohort drifts back toward average next period — which is exactly what the selected firms do (forward excess ≈ 0).

Why does the post-2001 record matter? Several "great" firms later failed badly (Circuit City, Fannie Mae). That forward collapse is what the no-skill null predicts and what durable skill would not — real-world corroboration of the artifact.

Is this just a simulation? Yes — deliberately the smallest one that isolates the mechanism, cloud-free and tuning-free, and a transparent runnable version of an already-published critique (Henderson–Raynor–Ahmed 2012; Denrell 2003/2004). The falsifier above states exactly what real forward-tested evidence would overturn it. Runnable code: research/probes/good_to_great_null.py (MIT, cloud-free) — re-run it or break it.