The order book as a lattice
An algorithmic trading operation streams four kinds of adjacency: the order book, where quotes link prices to sizes and update at the millisecond; news feeds, which link entities to events; filings, which link firms to obligations and ownership; and cross-asset signals, which link one market's move to another's expected move. None of these streams is interesting on its own for the question this page asks. What matters is whether they are connected — whether a shock in one corner of the graph has a path to somewhere a position sits.
Percolation theory was built for exactly this kind of question, though not on this kind of graph. Simon Broadbent and John Hammersley set it out in 1957 to understand gas moving through the granular carbon of a coal miner's respirator: not how the gas diffuses, which was already understood, but how the medium's own randomness governs whether a path exists at all. Harry Kesten proved in 1980 that bond percolation on the square lattice has a threshold of exactly one half. Below that density of open bonds, the filter blocks. Above it, gas crosses almost surely, and the transition between the two states is sharp, not graded. A filter that degrades two percent in porosity does not get two percent worse. It stops working, at one specific density, and not before.
A systematic PM lives inside a version of that filter. A signal that correlates two instruments is a bond in a graph. A cross-asset arbitrage strategy is a bet that certain bonds stay closed — that the correlation structure it depends on remains disconnected from noise, or connected in the specific pattern the model assumes. The question that matters for that PM's book is not "how strong is the correlation on average this quarter." It is whether the graph of dependencies the strategy relies on has quietly crossed a threshold: gone from a set of small, isolated clusters of co-movement into one spanning component where everything now moves with everything else, usually right before a liquidity event forces it.
Two positions
Position one. Markets are not lattices. Real financial networks are scale-free — a small number of hub instruments (index futures, the ten-year, a handful of mega-cap names) carry a disproportionate share of the connecting edges. In networks like that, the classical percolation threshold often vanishes: a spanning cluster can exist at almost any density, because the hubs alone are enough to connect everything. If there is no sharp p_c to cross, the whole apparatus of "detecting the threshold" is a physicist's fantasy imported into a market structure that doesn't have one.
Position two. That is true, and it does not help. A vanishing threshold does not mean connectivity is insensitive to change; it means connectivity is now governed by which few edges are open, not how many. Remove or add exposure at a hub — a single large fund unwinding a cross-asset carry trade, a central bank altering the correlation between rates and equities — and the giant component reorganises with no proportional warning in aggregate volume or average correlation. Heterogeneity does not restore the comfortable, smooth picture where a strategy degrades gradually as the world drifts. It makes the crossing depend on individual edges a static snapshot cannot distinguish from noise. This is worse for the frozen-corpus problem, not better.
Both positions are correct about the mechanism. They disagree about what follows. Position one is right that the textbook lattice threshold — one clean number, one clean transition — does not transfer literally to an equity-rates-credit-commodity graph with fat-tailed degree distribution. Position two is right that this makes the PM's monitoring problem harder rather than obsolete: the crossing is real, it is just edge-specific instead of density-specific, and edge-specific crossings are exactly the kind of event a periodic review is built to miss, because nobody reviews "did this one hub edge change" on a fixed calendar.
Where the failure actually happens
The characteristic failure in this domain is not a strategy that is wrong. It is a strategy that was right and is now trading a graph it no longer belongs to. A signal decays silently. It kept generating the same expected edge in backtest, kept passing the same significance tests on a rolling window, and kept executing — because the sample window used to validate it still straddles both the pre-crossing and post-crossing regime, diluting the change into something that looks like ordinary noise rather than a structural break.
This is the same shape as the 2003 Northeast blackout, the 2021 Suez closure, and the 2011 Thai flood disruption to chip supply chains: in each case a margin of safety was computed against a recorded topology, and the actual topology had already moved. A systematic PM's risk limits are the trading-desk equivalent of those margins. VaR computed on a correlation matrix estimated over the last sixty days is a belief about current adjacency. It is not a law. When a bond between two previously weakly-linked asset classes closes — say, a stress episode that couples credit spreads to a commodity that used to trade independently — the correlation matrix used for sizing is instantly a description of a graph that no longer exists, and it will say so nowhere in its own numbers. It just quietly prices risk for the wrong world.
Why retraining on a schedule does not fix this
The instinct is to retrain more often: recompute the correlation matrix daily instead of weekly, refit the signal monthly instead of quarterly. This assumes error accumulates gradually between refreshes, so a shorter interval buys proportionally less exposure to staleness. Percolation breaks that assumption specifically. Near a threshold, the correlation length of the system diverges — small local changes propagate arbitrarily far — and a graph can sit visibly stable for a long stretch and then reorganise across a handful of trading days. The crossing does not respect a retrain calendar, because it isn't a function of elapsed time. It's a function of accumulated adjacency, and adjacency arrives in bursts: a merger announcement, a central bank pivot, a single large fund's forced deleveraging that reconnects two previously separate liquidity pools.
A Large Language Model, in this frame, is the correlation matrix computed once and trusted afterward — a picture of the graph's density at the cutoff, presented with the confidence of settled fact even after the world has moved. A Large World Model is the live order book feed: excellent local resolution, tick by tick, but a spanning cluster is a global property of the whole cross-asset graph, and no amount of clarity about the room you're in tells you whether the corridor connecting it to the next room just opened. The position this page is arguing for — every stream held open, belief about global connectivity kept revisable, provenance retained on which edge moved the estimate — is the only regime that has any chance of catching the crossing near the moment it happens rather than in the postmortem.
What continuous intake actually buys, and what it doesn't
It does not buy omniscience. Detecting a percolation transition requires some global view of the graph, and no PM's system has that any more than a physicist watching one lattice realisation does. Streaming order books, filings and cross-asset feeds continuously produces a statistical estimate of connectivity, not a certainty, and that estimate is still built from finite, noisy samples of edge arrivals.
What changes is the shape of the error. A frozen snapshot gives one sample of adjacency and no way to estimate its derivative — you cannot tell, from a single correlation matrix, whether the graph is drifting toward a crossing or sitting comfortably subcritical. Continuous intake gives you the arrival process itself. Finite-size-scaling-style estimators — the distribution of co-movement cluster sizes across the book, a susceptibility measure that spikes as a spanning cluster nears formation — sharpen as edges keep accumulating, and they can flag the approach to a crossing days before the average correlation metric moves at all. You still find out late relative to the crossing itself. You find out in days rather than at the next quarterly review, and with provenance intact you can say which instrument's edge caused the estimate to move, which is the difference between a PM cutting a position on a hypothesis and cutting it on a guess.
The narrowed claim
Not every risk in a trading book is a connectivity belief. Funding cost, fee schedules, a single instrument's realised volatility — these vary smoothly enough that periodic review is adequate and cheaper than continuous monitoring. The claim that survives the two objections above is narrower than "watch everything constantly": it is that whichever beliefs in the book depend on the global structure of a changing dependency graph — is this pair still weakly coupled, has this hub reconnected two clusters that used to trade independently — are structurally invisible to sampling, no matter how heterogeneous or lattice-unlike the graph turns out to be. Identify those beliefs specifically. Those are the ones that force the stream to stay open, and the PM's responsibility is knowing which ones they are before the graph tells them the hard way.