The Disagreement

TAM-ARB.27 · Arbitrage · The Approximate Mind

Two households with identical finances, in 2015, asking two different advisors the same question, received different answers.

One was told to refinance now; the other, to wait for the next cut. One was moved into international equities; the other was kept home. One was told the emergency fund comes first; the other, that the match comes first and the fund can follow. The advice differed because the advisors differed, in training, in incentive, in temperament, in which decade formed them, and the households, following different advice, did different things at different times. Nobody designed this. Disagreement among advisors was a defect of the profession by every standard the profession applied to itself, a scandal to its academics and an embarrassment to its credentialing bodies, and it had one property nobody priced: it decorrelated the behavior of millions of similar households.

The defect was a stabilizer. When ten million households face the same rate environment and act on staggered, contradictory guidance, their refinancings smear across quarters, their reallocations offset, their exits stagger. The system absorbed shocks partly because its participants could not agree on what to do about them.

Models converge. Trained on overlapping corpora toward overlapping objectives, benchmarked against each other, tuned on the same published consensus of sound household finance, the optimizing instruments give similar households similar answers at similar moments, and the answers execute by default, as the taken-advice essay established. The claim: convergent advice executed at household scale is the coherent-crash mechanism running through ten million ledgers instead of ten thousand funds: the same refinance week, the same rotation out of a sector, the same exit from a product class, arriving as one correlated event that no individual ledger did anything wrong to produce.

The machine, in one sentence: optimization replaces a population of disagreeing human advisors with a small number of converging instruments, and the dispersion of household behavior, an unpriced stabilizer nobody built, is removed from the system without anyone deciding to remove it.

The Declared Scenario

#

The premise above is asserted, not measured. No empirical study exists comparing the dispersion of human advisory recommendations against the dispersion of model recommendations for matched households, and as far as the record shows, nobody is running one. What follows is therefore openly a hypothetical, and wears the label rather than burying it: everything below prices a scenario that has been declared, not demonstrated.

The study that would settle it can be named precisely, which is part of the point of the declaration. Take a standardized panel of household profiles, identical in assets, debts, income, and horizon. Put each profile to a broad sample of human advisors and record the recommendation set. Put the same profiles to the deployed household instruments and record theirs. Compare the dispersion of the two distributions, on timing as well as allocation, since the correlation risk lives in the when at least as much as the what. The study is cheap by the standards of the risk it would size, it requires no proprietary data a regulator could not compel, and its absence after several years of deployed household optimization is itself a small finding about where the field points its instruments.

Why declare before anyone measures? Because requiring empirical proof before a risk may be named is a suppression mechanism: unstudied risks stay unstudied precisely because nobody surfaced them, and the surfacing has to precede the study or the study is never commissioned. The correct instrument for an unmeasured structural risk is not silence and not assertion dressed as measurement. It is the declared scenario, priced conditionally, dated, and falsifiable.

The claim, dated. Horizon: end of 2030. The claim: by then, either a dispersion study of the shape named above will exist and will show model-driven recommendation dispersion for matched households materially below the human advisory baseline, or a correlated household-scale event, a synchronized refinancing wave, product exit, or reallocation traceable to convergent instrument guidance, will have occurred and been documented. Miss condition: if by 2031 the study exists and shows dispersion comparable to the human baseline, whether because instruments differentiate more than assumed, personalize more than assumed, or randomize deliberately, then the disagreement premise fails, the accidental-stabilizer finding stays historical, and the claim stands on the record as a scenario that did not arrive.

What the Scenario Prices

#

Conditional on the premise, three consequences sort out, and they land on different parties.

For the system, the removal of behavioral dispersion converts idiosyncratic noise into systematic exposure. The individual household’s optimization is locally correct in every case; the aggregate of locally correct, synchronized adjustments is a new class of event with no responsible party, since no instrument malfunctioned and no household erred. The coherent crash the series priced in professional fund management, correct models converging until their correctness became the event, needed only ten thousand funds. Ten million ledgers converging is the same mechanism with a wider base and a thinner set of circuit breakers, because household finance has no trading halt.

For the counterparties, synchronization is legible and therefore harvestable. A refinancing wave whose timing is predictable from the public behavior of the dominant instruments is a wave somebody positions ahead of, and the spread extracted from anticipating optimized households is a new arbitrage created by the closing of the old ones, fully inside this series’ oldest pattern: the audit does not end the game, it moves the table.

For the instrument builders, dispersion becomes a design variable with no owner. Deliberate de-correlation, randomized timing inside a tolerance band, staggered execution windows, heterogeneous model portfolios, is technically trivial and commercially unrewarded, since every household wants the best answer now and no household pays for the system’s stability. Whether any deployed instrument randomizes, and whether any regulator has asked, are empirical questions the declared scenario leaves standing where the study would find them.

The Frame

#

Decision rule: treat behavioral dispersion as an unpriced asset of the pre-optimization economy, assume it is being withdrawn at the pace of enrollment, and hold nothing whose solvency assumes household behavior remains staggered, until the dispersion study exists and says otherwise.

The advisors were wrong in different directions, and the system stood partly on the differences. The instruments will be right in the same direction, and nobody has measured what stands on that.

The Approximate Mind is a series exploring what AI actually does to human life. Arbitrage prices what the audit does to every spread it reaches.