A position held against a lifted constraint
A desk quant at a power trading firm builds a spread position on the assumption that a transmission constraint between two zones will bind through the week — congestion that has held for eleven straight days, priced into every forward curve on the desk. The constraint is a maintenance outage on a 500kV line, filed with the regional transmission organisation three weeks earlier. Overnight, the utility completes the repair ahead of schedule and files an update. The constraint lifts. The desk quant's model, still running on the cached outage notice, holds the position exactly as it was: long the congested-side hub, short the relieved side. By the time the change propagates through the model's next scheduled refresh, the spread has collapsed and the desk has taken the loss on a fact that was public, filed, and simply not yet read.
This is not a failure of intelligence. The quant is skilled, the model is well specified, the historical calibration is sound. It is a failure of intake — a belief held past the moment its supporting evidence expired, because nothing in the pipeline was watching the filing that revoked it. Deliberate practice, the cognitive-science construct behind decades of expertise research, gives a precise language for why this failure is not incidental to trading skill but definitional of it.
The construct, and what it actually says
Anders Ericsson, with Ralf Krampe and Clemens Tesch-Römer, published the founding study in 1993 on violinists at the Berlin Hochschule der Künste. Hours of accumulated experience explained skill weakly; hours of a specific kind of effort — tasks chosen just past current ability, performed with full attention, scored immediately against a correct standard, then repeated with the error fixed — explained it well. The finding relocated expertise from a property of the person to a property of the training loop. Exposure accumulates. Practice corrects. The two are not the same activity wearing different names, and mistaking one for the other is the single most common error in how skill gets discussed.
Applied to trading, the distinction is not abstract. A desk quant who has watched fifteen years of price data has exposure on a scale most professions never reach. Whether that exposure produced calibrated judgement is a separate question entirely, and it turns on how tightly each forecast was scored against what actually happened, and how fast.
Position one: continuous intake is what makes correction possible at all
The strong case for open, live streams — grid telemetry updating every four seconds, outage notices filed continuously, weather reanalysis reissued on each model run, regulatory filings arriving as they are docketed — is that without them, deliberate practice cannot occur in this domain at all. A model trained once on a frozen slice of grid history has exposure to how congestion patterns looked in that slice. It has no way of learning that a specific constraint, filed on a specific date, no longer holds, because nothing tells it the constraint changed. This is structurally identical to the anaesthetist before pulse oximetry, inferring oxygenation from the patient's colour and finding out about failure at the postmortem: the frozen record contains everything except the one signal that would let you correct in time. Continuous intake collapses that latency. A live feed of transmission filings turns "the constraint lifted three weeks ago and nobody told the model" into "the constraint lifted at 02:14 and the position closed at 02:19."
Position two: more streams, more noise, not more calibration
The counter-case does not deny the mechanism; it denies that the desk's actual streams supply it. Grid telemetry is high-frequency and largely irrelevant to the specific forecast error that sank the position — megawatt flow readings do not announce that a constraint's legal status changed, only that flow patterns shifted, which is a lagging and ambiguous signal. Weather reanalysis reissues introduce their own noise: each run revises the prior run, and a desk quant who re-optimises against every revision is not practising, in Ericsson's sense, at all. Ericsson's own protocol insisted on a coach — someone who chooses which specific error to correct and at what difficulty. A trading desk drowning in four-second telemetry, outage notices, weather reissues and regulatory dockets has volume without curation, and volume without curation is exactly the profile of the twenty-year clinician who plateaus rather than improves. More data streaming in does not, by itself, tell the quant which of last month's forty positions to re-examine and why one specific one went wrong.
The desk doesn't need more feeds. It needs someone — or something — to say: this position, this assumption, this is the one that was wrong, and here is the fact that made it wrong. Streams without that are just more noise arriving faster.
Where the objections land
Both objections above deserve a direct answer, not a dismissal.
The effect-size critique — Macnamara's 2014 meta-analysis found deliberate practice explains roughly 4% of performance variance in professions, far less than Ericsson's early claims implied — is real and applies here. Trading skill is not reducible to feedback tightness; risk appetite, capital constraints, and raw pattern recognition built over a career all matter, and none of them are supplied by better intake. But the claim this page is defending is narrower than sufficiency. It is that feedback is necessary — that no volume of exposure without outcome information produces calibrated forecasts — and that claim survives Macnamara intact. Her low figure for professions is itself evidence for the mechanism: professions are exactly where feedback arrives late and gets misattributed, which is why a desk quant's forecasting record can look, from a distance, indistinguishable from a novice's despite fifteen years on the desk. The 4% is a measurement of how badly most professional feedback loops are built, not proof that fixing them is worthless.
The curation objection is sharper and should be conceded almost in full. Continuous intake is not deliberate practice. It is the raw material deliberate practice requires and cannot exist without. A four-second telemetry feed does not select the task at the edge of the quant's competence, and it does not identify which specific belief to re-score. But the coaching function — deciding which error matters, which position to re-examine, which constraint to re-check — is itself a function over the feedback record. You cannot choose which of forty positions was wrong without knowing, in near-real-time, which forty positions exist and what became of each. The frozen corpus cannot supply that function at all, because it has no record of this desk's own recent calls. Continuous intake does not replace the coach. It is the precondition for the coach existing.
The three positions, restated for a trading desk
| what it has | what it cannot see | |
|---|---|---|
| Large Language Model | historical price series, filed tariffs, past outage records, read once at a training cutoff | any outage filed, or lifted, after the cutoff — the desk quant's own trade last Tuesday does not exist for it |
| Large World Model | live telemetry within a bounded session — flows, prices, dispatch at this moment | consequences that mature past the session: a constraint filed today that resolves in March, a regulatory challenge that plays out over quarters |
| Large Universe Model | grid telemetry, outage notices, weather reanalysis and regulatory filings held as open, revisable, provenanced streams across the full life of a position | nothing that is still reporting — the deficit becomes attribution and trust, not absence of signal |
What narrows, not what vindicates
The lineage argument survives this exchange, but smaller than it started. It is not that continuous intake makes a trading desk expert; Macnamara's numbers stand, and raw talent, risk discipline and market structure explain more of a trader's edge than any feed ever will. It is not that open streams are self-curating; the coaching objection is right that volume without a difficulty-matched, error-specific correction signal produces plateau, not mastery — and a desk that treats every telemetry tick as a lesson will learn less than one that scores its calls weekly against outcomes. What survives is the narrower and more defensible claim: correction requires an outcome to correct against, that outcome must be attributable to the specific belief that produced it, and a frozen corpus or a bounded scene cannot supply either. Grid telemetry, outage notices, weather reanalysis and regulatory filings, held open and provenanced past the life of the position, are the precondition without which the desk quant's fifteen years remain exposure. Whether they become practice is still, stubbornly, a matter of discipline the streams cannot provide by themselves.