The Lagging Truth

The Divergence Theorem for Regime Transition Detection — the plain-English companion

the paper (DOI) · code & data repository

A short explainer covering the same ground as the article below. The written companion carries the full detail.

Companion to the research paper of the same title. This is education, not investment advice. Nothing here tells you what to buy, sell, or predict. It explains what the paper found, how the checking worked, and what the results do and do not mean.

The claim in one sentence

The fast-minus-slow trick that traders call MACD and economists call the Sahm Rule is one mathematical object — this paper proves twenty-two results about it, tests it across seven sciences, and publishes the mixed scorecard as it fell: five wins, zero contradictions, two misses, finance among the misses.

The idea: one trick with many names

Across wildly different fields, people keep reinventing the same move. Take a fast, recent estimate of how much something is churning. Take a slower, longer-memory estimate of the same thing. Subtract. When the gap widens, suspect that the system’s regime is changing.

Finance calls this MACD. Macroeconomics calls a version of it the Sahm Rule — the recession detector built on fast versus slow unemployment averages. Process engineers call theirs CUSUM. Ecologists watch rising variance for ecosystem tipping points; neuroscientists build seizure predictors the same way. Each field treats its version as a homegrown heuristic. This paper treats them all as instances of one construct — the divergence operator: a fast, short-memory average of something, minus a slow, long-memory average of the same thing. Then it asks what mathematics can say about that object in general.

The answer is: quite a lot. Twenty-two proved results — theorems, propositions, corollaries, lemmas, all with complete proofs in the appendix — characterize the operator’s sensitivity, its timing, and its best configuration near a tipping point. Three deserve plain-English translation.

You can know the direction of the prediction before you look. The Persistence-Sign Theorem says the sign of the gap’s forward prediction is fixed by a single measurable property of the data: its persistence. That’s whether the series tends to keep doing what it was just doing (like market turbulence, which arrives in streaks) or to snap back (like some natural cycles). Persistent data: a widening gap predicts more of the same ahead. Anti-persistent data: the same widening gap predicts the opposite. One number, measurable in advance, tells you which way the needle points. That converts a folk indicator into an instrument with a spec sheet.

The window sizes are a dial with a law behind it. The theory derives the best fast window near a transition — about half the slow window — and a conservation law with the flavor of a physics result. You can tune the operator to fire earlier, but only by paying sensitivity, and the product of speed and sensitivity is fixed. There is no free lunch in the windows, and now that’s a theorem rather than a trader’s hunch.

The reigning theory is a special case. The dominant scientific framework for early warnings is critical slowing down: near certain tipping points, systems recover from small shocks more slowly — their wobbles grow bigger, and each wobble starts to resemble the one before it. The paper proves that critical slowing down is what the divergence operator becomes when you restrict it to that one class of transition. The operator carries strictly more information — which also means it can watch for regime changes that critical slowing down is structurally blind to.

What we tested, and what happened

One specification — windows, transforms, pass/fail gates — was frozen and committed before any analysis ran; the repository’s first-commit timestamp is the pre-registration. It was then applied unchanged across seven domains: equity volatility (how violently stocks swing), equity correlation structure (how much stocks move together), river hydrology, solar physics, influenza surveillance, regional weather, and macroeconomics. Every input file is fingerprinted; every number in the results section is rendered from a machine-checked ledger rather than typed.

The verdict is mixed, and the paper leads with that word. The sign law — the Persistence-Sign Theorem’s directional prediction — holds in five of the seven domains and is contradicted in none; the two remaining domains are misses rather than reversals. Solar physics is the strongest confirmation. Influenza surveillance confirms strongly too — the operator’s short-window forecast shows the strongest link to the future in the whole panel. Equity correlation structure confirms modestly. And the practical question — does divergence actually beat critical slowing down where both apply? — gets a real but domain-dependent yes: an edge in some fields, parity in others, no universal victory.

Then there is the paper’s weakest result: the headline equity-volatility panel — the finance case, the one closest to the MACD folklore that motivates the whole subject. Its average link between signal and outcome across the panel is +0.05 — on a scale where 0 means no link at all and 1 means perfect lockstep — and it failed the pass/fail gate it set for itself in advance. The abstract says this in plain terms. The verification suite regenerates the failure on demand, gate verdict and all, every time anyone runs it. A framework built to explain fast-versus-slow indicators, reporting that its flagship financial test didn’t clear the bar it set for itself, is the strongest evidence in the paper that the other five domains’ confirmations mean something.

The one out-of-sample check

Everything above is in-sample: the tests and the data were both on the table when the paper was built. The paper reports exactly one clean exception, and it is careful about what to call it. The solar specification — including the operator’s negative predicted sign, called in advance, an unusual and risky call — was fixed on solar records ending in 2008, before any of the later data existed to influence it. On the post-2008 solar data — data the specification never saw — the negative sign persists. And the paper states that it is predicted to continue holding on solar data published after the paper — checkable by anyone, indefinitely, against a specification that cannot now be revised.

The paper is explicit about what this is and isn’t: a verification of a locked operator, not a promoted live forecast. No dated public prediction with a kill switch rides on this paper — that program belongs to other papers in this research line. What this paper offers is narrower and clean: a sign called in advance, surviving seventeen years of data that arrived after the call, with the call’s paperwork committed. The held-out stretch is short — roughly one solar cycle of annual readings — so the sign, rather than the size of the effect, is the operative test, and the paper says so itself.

What the checking caught in this paper

The general machinery every paper in this series runs through — the hash-pinned inputs, the machine-checked ledger of numbers, the verification program, the adversarial review — is described once in the series’ shared verification note, which follows every companion on LaggingTruth.com. What belongs here is what the process caught and changed in this paper specifically.

The theorems were checked three independent ways: by hand in the complete written proofs, by a numeric stress test on small samples, and by a computer program that redoes the algebra symbolically and checks every sign. The latter two run from a committed script, and a second committed script reconciles all three against each other, so a proof cannot silently drift from what the machine verified. The adversarial review — a fresh, memory-isolated AI session under a fix-or-rebut protocol — returned one load-bearing finding: the paper had mis-reported which domains constituted the sign law’s five confirmations. The count of five-of-seven was right; the membership list wasn’t. It was corrected, with no proof and no reproduced number affected. The full record — prompt, transcript, and the disposition of that finding alongside nine minor ones, each fixed or rebutted with a written reason — is committed verbatim in the repository. Citations were verified in tiers: existence and accuracy for every reference, claim-support for the ones the argument leans on, full-text confirmation for the load-bearing few.

The equity failure also belongs in this section: the gate was pre-registered, the gate tripped, and the verification suite preserves the tripped verdict — anyone who re-runs it regenerates the failure, gate verdict and all.

What this cannot do

  • It is not a universal alarm. The verdict is mixed by the paper’s own account: five confirmations, two misses, and a practical edge over the incumbent theory that depends on the domain. Nothing here licenses pointing the operator at an arbitrary system and trusting the output.
  • Finance is its weakest domain. Read that again, because the subject invites the opposite assumption: the equity-volatility panel failed its own pre-written gate. Anyone hoping this paper validates MACD-style trading signals will find it reports the closest thing to a refutation it was empowered to produce.
  • The direction requires a measurement first. The Persistence-Sign Theorem’s power — knowing which way the prediction points — is conditional on measuring the data’s persistence correctly. Get that wrong and the sign flips on you.
  • The window dial obeys a conservation law. Earlier warnings cost sensitivity, provably. Any configuration is a trade, not an upgrade.
  • The solar result is a verification, not a forecast product. The paper checks a specification locked on data ending in 2008 against later data and predicts the sign persists on future data — a falsifiable statement anyone can monitor. But it deliberately promotes no dated forecast and attaches no decision rule to it.

The claims stay attackable: the pre-registered specification, the fingerprinted data, the ledger, and the gate that the equity panel failed are all in the public repository. And the solar sign prediction stands exposed to every year of data the sun has yet to publish.

The takeaway

Humanity keeps rediscovering the same early-warning trick and naming it after whichever field found it last. This paper gives the trick its mathematics: twenty-two proofs, a theorem that reads the prediction’s direction off the data’s own persistence, a law governing what the windows can and cannot buy, and the incumbent theory recovered as a special case. Then it does the harder thing — tests one frozen specification across seven sciences and publishes the scorecard as it fell. Five confirmations, zero contradictions, two misses — with the finance flagship failing its own gate, and one solar call, locked on data ending in 2008, still standing after seventeen years of data that arrived afterward. The gap between fast and slow is real information with real limits, and for the first time both the information and the limits come with proofs.

This companion is licensed CC BY-NC 4.0. The research paper it accompanies is licensed CC BY-NC-ND 4.0, and the analysis and verification code is MIT-licensed. Education, not advice: nothing in this document is financial advice, an investment recommendation, or a forecast.