The objection that should win
Here is the case against this whole framework, stated as a systematic portfolio manager would state it. Markets are not organisms. They have no set point. Nobody is defending glucose or core temperature; there is no evolved variable that a trading system forecasts because failing to forecast it would kill the organism. Allostasis describes bodies with skin in the game — literal, physiological stakes. A trading book has positions, not stakes in that sense. To borrow a term built for cephalic-phase insulin release and apply it to a signal decaying in an order book is to dress up ordinary predictive modelling in biological costume. Quant desks have run anticipatory models — volatility forecasting, regime detection, ahead-of-print filing sentiment — since well before anyone on the desk had heard of Peter Sterling. Calling it allostasis adds a metaphor, not a mechanism.
That objection is close to correct about the label, and this piece will not fight it on that ground. It is wrong about what follows.
What the desk actually does all day
A systematic PM does not react to the market. Reaction is what stop-losses do, and stop-losses are the least intelligent instrument on the desk, kept precisely because forecasts fail. The rest of the operation is anticipatory. Order book imbalance at the top five levels predicts short-horizon price pressure before it shows up in trades. A filing feed predicts an earnings surprise before the print. Cross-asset signals — credit spreads widening ahead of an equity drawdown, dollar funding stress ahead of an emerging-market currency move — predict a regime change before the position-level P&L shows it. None of this is correction after the fact. It is a forecast, built from a running history of the same signal's past behaviour, positioned ahead of the event it concerns.
And every one of those forecasts is continuously re-scored. A momentum signal that worked in 2021 is tracked against its own realised hit rate this month, this week, today. The PM does not wait for a drawdown to ask whether the signal still works. The question is live, permanently, because the streams that feed the answer never stop arriving. That is the mechanism this piece cares about, independent of what it is called. Sterling and Eyer's point about blood pressure was never really about blood pressure. It was that a defended variable can be moved ahead of demand only by a system still taking in the evidence the demand is built from. Substitute "spread" for "pressure," "order flow" for "arousal," and the structure holds without needing the word allostasis to do any work.
The failure that names the whole domain
The characteristic failure on a systematic desk is not a bad forecast. Bad forecasts are priced in; every strategy has a known error budget. The characteristic failure is a signal that decays silently and keeps trading after it has stopped working. A factor built on a data source that has been arbitraged away by competitors, or a news-sentiment model reading a feed whose composition has shifted upstream, continues to generate positions with the same confidence it had when it worked. Nothing in the system's output changes shape. The P&L erosion is gradual enough to hide inside ordinary variance for weeks. This is not a control problem in the classical sense — there was no single error to correct. It is a provenance problem: the forecast is still being made, but nobody can any longer trace it to evidence still arriving that supports it.
The Large Language Model version of this desk would be a model trained once on a historical corpus of order books and filings, its weights frozen, its view of "how order flow predicts price" fixed at the training cutoff. It cannot even suffer silent decay in the interesting sense — it never had intake to lose. Retraining is the only repair, and retraining happens outside the trading day, on somebody else's schedule. The Large World Model version is the intraday execution algorithm: it senses the order book, predicts the next few ticks, acts, re-senses, and does this well, but only across the life of the parent order. Ask it whether the signal it is executing on still works in a structural sense and it has no memory long enough to answer. The Large Universe Model condition is the one the PM is actually trying to build toward and never quite reaches: every stream — order books, news, filings, cross-asset — still running, every belief about a signal's continued validity held with a record of exactly which evidence produced it, so that when the evidence stops supporting the belief, the belief can be retired rather than discovered dead in a drawdown report.
Objection: intake without an objective is not control
A fair challenge follows immediately. A desk fed every tick, every filing, every cross-asset print in existence has no principle for deciding which of those deviations are worth a position. Continuous observation is not a strategy. Evolution gave the organism glucose and core temperature to defend; nobody hands a trading system a defended variable. Feed it everything and it will find spurious correlation in all of it, which is a worse failure than decay, because it trades confidently on noise from day one rather than decaying gradually from a real edge.
This is correct, and the claim here is narrower than it sounds. The axis under discussion is intake, not mandate. A firehose of every market stream regulates nothing until a PM specifies what the book is actually defending — a target volatility, a factor exposure, a Sharpe ratio net of costs. Objectives decide what is worth forecasting. Intake decides whether the forecast can be made and revised at all. A PM with a sharply specified mandate and only end-of-day data is bounded above by that data no matter how well-specified the mandate is; a PM with full continuous intake and a sloppy mandate will simply lose money faster. The two failures are independent. The point stands only where the objective is fixed: given a stated signal, the quality of its anticipatory control is capped by what is still being observed about it and for how long. That cap is what terminates on this axis, not the wisdom of the mandate.
Objection: more streams, worse control
The harder challenge is the one from allostatic load. Bruce McEwen's extension of the theory was precisely that chronic anticipatory adjustment has a cost, and the cost can be the regulation itself turning pathological — sustained cortisol elevation degrading the very tissue meant to benefit from foresight. The trading-desk analogue is well known and has a name: overfitting to live data, or more specifically, a risk system that has ingested so many correlated cross-asset signals that it forecasts a regime break in almost everything, freezing the book in permanent defensive crouch. A PM who wires every feed into the risk model without discipline does not get better anticipation. Gets a system with a fever, seeing danger in every tick because it has stopped being able to say which of its thousand inputs actually earned its alarm.
That risk is real and it is not a rare failure mode; it is close to the modal one on desks that add data sources faster than they add discipline. But look at where the damage sits. It sits in the weighting and the decay function, not in the fact of intake itself. A hypervigilant risk model is not suffering from having too many streams; it is suffering from having lost the provenance that would let it say "this alarm traces to this evidence, and that evidence is now stale, so retire the alarm." Give the same intake a proper record of which signal justified which position, timestamped and decayed on a schedule matched to the signal's own half-life, and the flood becomes tractable. The failure McEwen documents is an argument for exactly the feature the terminal position is built on — provenance sufficient to retire a belief — not an argument against continuous intake.
What the desk can actually claim
The narrower claim, then. A systematic PM cannot forecast a signal's decay from a frozen corpus, because decay is precisely the thing that happens after the corpus was fixed. A PM cannot forecast it from within a single execution episode either, because decay is measured in weeks and the episode is measured in minutes. Anticipatory control of "does this signal still work" requires streams that keep arriving after the position is put on, and a record good enough to trace a bad forecast back to the evidence that once supported it. Once a desk has that — everything reachable, still arriving, with provenance — there is no further category of market evidence being withheld from it. There is more of it, better weighted, held longer. Those are matters of engineering and discipline, not a new rung on the ladder.