The loop, from tick to position
A desk quant trading power and gas exposure sits above four live streams. Grid telemetry arrives every four seconds from the transmission system operator: frequency, interconnector flow, constraint status. Outage notices arrive irregularly, sometimes seconds before a unit trips, sometimes days ahead as planned maintenance. Weather reanalysis refreshes every six hours, revising the wind and temperature forecasts that drive demand and renewable output. Regulatory filings — a change to a capacity mechanism, a new curtailment rule, an emergency instrument from the energy regulator — arrive on no schedule at all, sometimes buried in a PDF nobody parses until the desk downstream notices prices behaving oddly.
None of this is a corpus. A Large Language Model trained on energy market documents would have internalised, at some cutoff, that a particular transmission constraint exists and binds in a particular way. That fact would sit in its weights as if permanent, because the corpus that produced it stopped moving the day it was collected. A Large World Model does better within a session: point it at today's order book and today's grid state and it will track deviations as the session runs. But the session ends, the position is closed or the shift changes, and the baseline it built evaporates. Tomorrow starts from zero gain again. Neither position holds a baseline that survives the boundary between one observation window and the next, and the desk's actual exposure spans weeks.
What is held
The quant's book does not hold raw ticks. It holds a running baseline for each constraint, each price relationship, each spread — the electricity-gas spread, the imbalance price against day-ahead, the interconnector flow against its historical range for this time of year. This is the trading equivalent of a photoreceptor's adapted state: not "the interconnector is flowing at 1,400 MW" but "1,400 MW is 0.3 standard deviations above what the last fortnight has taught us to expect, given the wind forecast". The absolute number is cheap to store. What is expensive, and what the desk actually depends on, is the current estimate of normal, and the memory of when normal last moved and why.
A position is built against this baseline. A short position on a regional imbalance price, for instance, might be justified by "this constraint has bound on 80% of days this month, so the price spikes it produces are structural, not noise". The position's whole economic logic depends on the constraint remaining in force. That dependency is the loop's weak point.
What triggers revision
Revision should be triggered by a regulatory filing lifting the constraint, or a system operator notice that the network reinforcement behind it has completed. In practice it is often triggered by something noisier: a settlement print that does not fit the adapted baseline, a spread that fails to reappear when the desk's model says it should. The rescaling event itself is not the arrival of new data — new data arrives constantly and mostly confirms the baseline. Revision is triggered when incoming telemetry diverges from the adapted expectation by more than the noise the desk has learned to tolerate, and that divergence persists past the point where it could be explained as a transient.
This is the crux of the failure mode. A fixed-gain reader — a rule fired off a threshold set months ago, or a model trained on a corpus that ends before the change — cannot register the divergence as meaningful, because it never had an adapted baseline to diverge from. A properly adapting reader registers it immediately, provided it is looking, and provided the filing that caused it was actually ingested rather than sitting unread in a compliance inbox.
The characteristic failure
The constraint is lifted overnight. A network reinforcement completes, or the regulator issues an emergency derogation suspending a curtailment rule, and the filing goes out at 11pm. The desk quant, who built the short position on the assumption that this constraint reliably binds, does not read regulatory filings at 11pm. The book carries the position into the next session unchanged. Telemetry the next morning shows the constraint not binding — flow through the interconnector at levels that should have triggered curtailment, prices flat where they should have spiked. On a properly adapting system this divergence would immediately widen the confidence interval around the position's assumptions and flag it for review. On the desk as actually run, the divergence is initially read as noise, because the trading model's baseline has not yet been told that the regime changed. The position is held for one, two, sometimes several sessions against a constraint that no longer exists. The loss is not from a wrong forecast. It is from a right forecast about a world that stopped being true.
What the operator sees, and what they should see
What the quant's screen typically shows is a P&L line and a handful of risk metrics computed against the desk's own historical volatility — itself a slow-moving baseline, often recalibrated weekly or monthly. What it should show, if the system were built on the terminal position of this axis, is a provenance trail alongside the position: the constraint's binding frequency over the trailing period, the date that frequency last shifted, and — critically — a live flag the moment telemetry diverges from the adapted baseline by more than the tolerance the position was sized against. The difference between these two screens is the difference between finding out about the lifted constraint from a bad settlement three days later and finding out from a flagged divergence three hours after the filing landed.
The cost of the gap is measurable. A short position sized at, say, 50 MW against a constraint priced at a typical spike premium of 30 GBP/MWh, held for three extra sessions after the constraint lifts, accumulates losses in the low tens of thousands before the desk notices via P&L rather than via evidence. The information that would have closed the gap — the filing — existed and was public from the first minute. The failure was not information scarcity. It was a reading system with the wrong kind of gain.
Objections worth taking seriously
Digital systems have no firing-rate ceiling. Store everything at full precision, rescale in post-processing whenever convenient. The retina's constraint is a hardware problem that trading infrastructure does not have.
This is true at the storage layer and false at the inference layer. A risk model fitted to eighteen months of history in which the constraint always bound has, in effect, a fixed gain even if every tick it was trained on is stored losslessly. Feed it tomorrow's unconstrained flow and it does not fail to store the number — it fails to know the number means something different now. Covariate shift is saturation without the excuse of limited bit depth. Post-hoc rescaling still needs someone to notice the shift and pull a fresh baseline, which is exactly the continuous-observation requirement restated, not removed.
A system that continuously rebases what counts as normal can no longer tell you what normal used to be — and a slow drift in constraint frequency, unlike an overnight lift, would be invisible to it. A frozen risk model, at least, gives you a fixed and auditable reference.
This lands. A pure adapting system is blind exactly to the gradual version of this failure: a constraint that binds 80% of the time in January and 40% by June, with no single filing to flag, is a slower version of the same problem and harder to catch by threshold alone. The answer is not to abandon adaptation but to timestamp it — retain the prior baseline, the date it moved, and the trigger, whether that trigger was a filing, a persistent divergence, or a scheduled quarterly re-estimation. A frozen model is auditable but wrong from the day the world moved. A model that rebases and logs each rebasing is auditable and current, and the log is precisely what lets a slow drift in constraint frequency become visible after the fact, as a series of small dated rescalings rather than one silent slide.
What continuous actually requires here
None of this requires a single running process rather than a well-designed batch job. A weekly re-estimation of constraint-binding frequency is a legitimate implementation of continuous intake if constraints change on a monthly timescale. It is not adequate for the overnight-lift failure, because the world in that case moved on a timescale of hours and the re-estimation interval must collapse towards that timescale or the gap simply reopens. The requirement is not continuous in the sense of always running. It is continuous in the sense of never being older than the fastest thing it is meant to catch — for grid telemetry, that is minutes; for regulatory filings, it is however long it takes someone, or something, to actually read them.