The Compound Trap — the plain-English companion

the paper (DOI) · code & data repository

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

A companion paper showed that banking’s early-warning gauge destabilizes the credit it regulates. This paper asks whether central banks, budgets, or markets can ride to the rescue — and finds, across five major economies, that none can, while pricing the two small fixes that work.

The car with a pull

The previous paper in this line, The Measurement Trap, ended with an uncomfortable result: the international rule that tells banks when to build rainy-day capital reads its trigger from a twenty-year trailing average of credit itself, and a feedback loop closed through that long a memory turns self-amplifying when the thing it regulates is persistent. Credit is among the most persistent series in economics. One loop, mis-wired.

But an economy is not one loop. It is more like a car with four coupled systems — and this paper builds that car. Alongside the credit/Basel loop it wires up three more: monetary policy (the central bank steering by smoothed inflation), fiscal policy (budget rules steering by estimated output), and financial conditions (markets tightening and loosening with stress). Each loop is parameterized from primary sources, the connections between them from estimated couplings, and the whole four-loop machine is assembled for five economies: the United States, the United Kingdom, Japan, the Euro Area, and South Korea.

The question is the natural one. If one wheel pulls, can the steering, the brakes, and the engine — working together — keep the car in its lane?

One number for the whole machine

For a coupled system like this, mathematics offers a single verdict number: the spectral radius, written ρ. Think of it as the machine’s amplification factor. Give the system a small shock and watch one round of all four loops respond: if ρ is below 1, the shock comes back smaller — the machine absorbs it, and repeated rounds grind any disturbance away. If ρ is above 1, the shock comes back larger, and rounds compound it. There is no middle: below one heals, above one grows.

The compound machine’s verdict: ρ above 1 in all five economies — 1.00998 for the United States, with the other four just behind. And the mathematics also says where the drift lives. Every unstable system has a dominant direction — the pattern the whole machine settles into as its errors grow — and here that direction loads almost entirely on the credit loop: over 99.7% of the weight, in every country. The steering, brakes, and engine are fine. The car drifts because of the one wheel, and the drift is the credit wheel’s drift, barely diluted by everything else attached to it.

Behind “all five” sit a construction fact and a measurement caveat, both stated in the paper. First, the strengths of the connections between loops were measured on US data and applied to all five economies, because comparable series for the others were not available at full quality. So how alike the five look is partly built in by construction; each one’s instability is not — it comes from its own credit data. Second, credit’s persistence was estimated on the level of credit, and a rising trend can push that estimate upward. The result is established for that measure and untested for detrended alternatives.

Two honest gradations belong beside that headline, and the paper prints both. The margin is thin — an amplification factor of 1.01 grows slowly, which is exactly why the problem is invisible in any quarterly report. And when the paper tested its pre-registered stronger claim — that credit’s dominance would sweep every diagnostic in its original strict form — the committed decision rules returned mixed verdicts on two supporting experiments. The paper reports them as mixed rather than rounding up.

Three rescues that fail

The paper then tries, systematically, to save the system without touching the credit filter.

Can the central bank offset it? The model’s answer is a search: how hard would monetary policy have to lean against credit to pull the compound ρ below 1? The search runs the counter-cyclical pass-through out to eleven times the strongest coupling observed in the data — pressing the brake eleven times harder than it has ever been pressed — and finds no stabilizing strength anywhere in that range. The instability barely notices: the credit loop’s persistence is so close to a pure unit root that other loops’ pushes dilute into it.

Does a crisis rewire things in our favor? One might hope that in crises, when couplings shift, the system self-corrects. The paper’s crisis-regime test finds the opposite lean: crisis-time couplings shift the compound system toward instability, not away — the machine is least stable exactly when policy is pushing hardest. Waiting for a crisis to teach the system stability is not a mechanism the estimates support.

Is it just the wrong kind of credit? Perhaps excluding mortgages, or consumer credit, would fix the gauge. The paper disaggregates US credit into five components — mortgages, corporate debt, consumer credit, business loans, commercial real estate — and finds every one individually unstable under the codified twenty-year window. The window, not the credit mix, is the binding constraint. Swapping tires does not fix a bent axle.

The fix that cannot work — proved

The natural repair, from this research line’s own toolkit, is the divergence element: replace the twenty-year gauge with the difference between a fast and a slow average, which reacts quickly and carries no long memory. Tested in isolation, it works — the credit loop alone stabilizes. Wired into the four-loop machine, it fails: across the entire pre-registered corridor of fast-window settings, the compound ρ stays above 1 in all five countries.

And here the paper does something stronger than reporting a failed experiment: it proves the failure is necessary. Theorem 1, machine-verified both symbolically and numerically, shows that any pure change-detector — any instrument built only from differences of averages — has exactly zero grip on the level of a series with a pure unit root. The instrument is deaf by construction to the thing that needs anchoring: like a smoke detector that responds only to changes in smoke, it will sleep soundly through a room already full of it, however sensitive you make it. No tuning, no window choice, no amplification can give a pure change-detector level force. The proof closes the reform menu’s most attractive door — including on the research line’s own favorite tool.

What actually works

Two repairs pass, and both are quantified. Shortening the gauge’s window from eighty quarters to twenty stabilizes all five countries (US ρ falls to about 0.983). Or, keeping change-detection, add a level anchor — a small dose of attention to where credit actually stands, not just how it is moving. The minimal stabilizing dose is 5 percent: ninety-five parts change-detection, five parts level, and every country’s compound machine goes stable (US at about 0.9998, with margin). The reform menu is closed and priced: shorten the memory, or add a pinch of level-awareness; pure change-detection is provably out.

The thermometer and the thermostat

Then comes the twist that reframes the whole debate — delivered by an experiment that went against this research line’s own candidate.

Defenders of the credit gap, including researchers at the institution that maintains it, have always had a strong card: the gap predicts crises well. The paper takes that claim seriously and runs a horse race on roughly 150 years of data across 18 countries: the official gap versus the divergence replacement, each scored on how well it ranks country-years by crisis risk. The gap wins, clearly — a score of 0.684 against the divergence’s 0.573, with the shortfall statistically resolved. The paper’s own preferred instrument is the inferior crisis sensor, and says so.

But the two results together dissolve the apparent contradiction, and this is the paper’s central principle. The gap is the better thermometer — and an unstable thermostat. Being the best measurement of where danger lies, and being a safe thing to wire to the control dial, are different properties, and the twenty-year gap has the first without the second. So the answer is not to smash the thermometer; it is to take it off the furnace switch. Measure with the best sensor; act through a stable control law. Keep the gap as the warning gauge it genuinely is; just don’t let a twenty-year memory steer the capital requirement mechanically.

Two bets at the next recession

The mechanism at the heart of all this — credit’s extreme persistence — makes two blunt, checkable predictions, and the paper registers both publicly, resolvable by anyone from named public data.

First: at the next US recession, the credit-to-GDP ratio will be higher two quarters after the business-cycle peak than at the peak itself. Not because lending booms in recessions — but because credit barely moves while GDP, the denominator, falls faster. The retrospective record is one-sided: +4.1 percentage points in 2001, +0.6 in 2008, +8.9 in 2020. Second: total US bank credit will not fall more than 3 percent from its pre-recession peak through two quarters past the trough. In the last three recessions the largest such decline was about a third of one percent — even 2008, the crisis synonymous with a “credit crunch,” saw aggregate bank credit dip only about two-tenths of a percent. The persistence isn’t a modeling assumption; it is the observable phenomenon.

Both come with exact data series and a step-by-step scoring protocol, and the first carries a ±0.10-point inconclusive band. Neither has a deadline: each waits for the next recession. If no dated US recession arrives before April 2031, they are recorded as untestable and carried forward — a test that never ran counts as nothing, not as a pass. Failures go to the public corrections log.

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 hostile review — a memory-isolated session reviewing a single curated package, re-running the verification gate before reading — returned one load-bearing finding, eleven material findings, and eleven minor ones. Every sustained finding was fixed rather than rebutted, and two fixes required entirely new committed experiments. The load-bearing finding was the sharpest: the paper’s model of the Basel trigger and the regulation’s literal wording admit two readings with no algebraic bridge between them. The remediation tested the conclusion under both readings — the instability holds either way, and the small level anchor repairs it either way. The same finding was exported to the companion paper, whose own Basel mapping inherited the issue and its remedy. The review also forced a sensitivity-closure analysis that now brackets the headline under coupling variation.

The pre-registered rules did the rest of the catching. Two supporting experiments returned mixed under their original strict decision rules and are reported as mixed, with one corrected reading logged as a dated amendment. The rolling historical check came back merely descriptive — two of its three pre-registered episode comparisons failed, and the failing legs are named in the paper with a power disclosure. And the crisis-sensor horse race delivered its verdict against the paper’s own replacement candidate, which is printed as the finding it is.

What this cannot do

• It is a model’s verdict, within its stated scope. The conclusion holds within the parameterization and the observed-to-stressed coupling ranges the paper commits to; it is not a claim about every conceivable policy response, and the paper’s limits section says so.
• The Greek episode is consistency, not proof. Greece’s post-2010 spiral is consistent with compound amplification under measurement-based rules, and the paper claims consistency only — that outcome had many causes.
• The gap keeps its warning job. Nothing here says the credit gap is useless — the paper’s own horse race shows it is the best crisis sensor tested. The claim is about wiring it to the control dial.
• Mixed verdicts stay mixed. Two supporting experiments did not clear their pre-registered bars in full and are reported that way.
• The predictions may never activate. Both wait on a dated US recession; by rule, no trigger before April 2031 means untestable and carried forward.

The takeaway

One mis-wired loop, it turns out, is not rescued by the other three. Built as a four-loop machine across five economies, the compound system amplifies in every one, and the drift belongs almost entirely to the credit loop. The rescues all fail — monetary policy at eleven times its observed strength, crisis-time rewiring, credit-mix surgery, and even this research line’s own divergence instrument, whose system-level failure the paper proves is unavoidable for any pure change-detector. What works is modest and priced: a shorter memory, or five parts in a hundred of level-awareness. And the horse race that crowned the official gap the better crisis predictor sharpened the conclusion into a principle any engineer would recognize — measure with the best sensor, act through a stable control law. Two public bets on the next recession now wait to grade the mechanism underneath it all.

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.