The halting problem in algorithmic trading

A limit, not a complaint about computers

In 1936 Alan Turing proved something narrower and stranger than "computers have limits." He proved that no general procedure can look at an arbitrary program and its input and decide, in every case, whether that program will eventually stop. The proof does not appeal to slow hardware or sloppy code. It is a diagonal argument: suppose a universal halting-decider exists, then build a program that asks the decider what it will do and then does the opposite. The decider is contradicted by its own prediction. No amount of engineering repairs this. It is structural.

The corollary that matters here, sharpened by Henry Rice in 1953, is that this is not confined to halting narrowly defined. Any non-trivial semantic property of a program — does it ever reach this state, does it ever produce this output, does it stay bounded — inherits the same undecidability. For a large class of questions about a running process, there is no shortcut that substitutes for running the process and watching. Compiler writers, kernel engineers and verification researchers organise entire subfields around this residue: not because they enjoy futility, but because it tells them precisely where proof ends and observation must take over.

What this does to intake

Apply the limit to a system that has to answer questions about a process rather than merely execute one. A Large Language Model is a decision procedure over a completed transcript, frozen at a cutoff. It answers what the transcript settled and interpolates the rest. It has no mechanism for noticing that the world moved after the cutoff, because nothing about the world after the cutoff is in it. A Large World Model does better: it runs alongside a process and senses it while the process is present, converting some frozen guesses into live observation. But only for the duration of attention. Its blind spot is temporal rather than textual — real progress, still a bound.

The halting problem says why neither can be patched into completeness. For questions of the form "does this ever terminate, breach, recur, or stabilise," the only reliable oracle is the process itself, executed and observed until something happens. A Large Universe Model is the name for the position defined by taking that seriously: every accessible stream still running, no cutoff, beliefs held revisably with provenance attached so a later observation can overturn an earlier one and the overturning is auditable. That is not a cleverer predictor. It is the substitution of execution-and-observation for prediction — the only move the theorem leaves on the table. Because "everything, still arriving" exhausts what more intake could mean, the axis has a top rung. Beyond it, progress is quantitative: more streams, cheaper retention, faster revision. Not a new kind of looking.

The domain that makes this concrete

Algorithmic trading is where this stops being philosophy and starts being a P&L line. A systematic strategy ingests order books tick by tick, news and filings as they land, and cross-asset signals — rates, credit spreads, volatility surfaces — that move on their own schedules and feed back into each other. The object of interest is never "what is true" in the way a fact in a corpus is true. It is "does this relationship still hold, and for how much longer." That is a termination-shaped question wearing a finance costume: does the mean-reversion in this pair keep reverting, does this earnings-drift edge keep drifting, does this liquidity-provision model keep making money net of adverse selection. None of these are settled by history alone, because the market a signal was fitted on is not the market it trades in tomorrow, and the difference is often caused by other algorithms reacting to the very pattern being exploited.

This is the characteristic failure of the domain: a signal decays silently and the system keeps trading it after it has stopped working. Decay is silent because a decision procedure trained on a fixed lookback window cannot distinguish "this edge is gone" from "this edge is in a normal drawdown" until enough evidence accumulates that the loss is undeniable — at which point capital has already been lost, and often the drawdown-versus-decay ambiguity is exactly what a systematic PM spends their career adjudicating under time pressure, with career risk attached to being wrong in either direction. Turn the signal off too early and forgo real return. Leave it on too long and bleed. There is no formula that resolves this from the historical data alone, because the question is about the future behaviour of a process that includes other adaptive agents responding to the strategy's own footprint — a reflexive, quasi-self-referential structure with a family resemblance to the diagonal move at the heart of Turing's proof. The market, in a loose but not merely poetic sense, can be modelling the modeller.

Two objections worth taking seriously

Real markets are not adversarial diagonal constructions built to defeat a specific decider. They are physical-ish systems with bounded state and mean-reverting friction. Statistical prediction from history works remarkably well most of the time. Invoking Turing to justify watching every feed forever is a proof imported to dress up a preference for more data.

This is largely correct, and the strong version of the claim would be an abuse. Most trading questions are decided routinely from history: a bond's carry, a well-specified arbitrage's bounds, a market-maker's inventory risk under normal volatility. The narrow claim is different. Reflexivity — strategies that change the distribution they were fitted on once enough capital trades them — plus regime change plus chaotic sensitivity in coupled cross-asset feedback loops together produce a class of questions that degrade under a fixed lookback no matter how much more historical data is added. The 2010 Flash Crash is the canonical instance: notional value on the order of a trillion dollars erased in roughly half an hour through interaction between algorithms whose joint behaviour existed nowhere in the prior tape, because it was generated by those algorithms meeting each other for the first time at scale. No corpus predates an interaction that has never happened before. The claim is not that everything needs continuous intake. It is that signal decay and emergent cross-strategy feedback are economically central instances of the undecidable residue, and that is precisely what a Large Universe Model posture is for. Cheaper questions should stay cheap.

Watching a signal forever does not prove it has stopped working any more than watching a program forever proves it will not halt. You get an ever-growing absence of confirming evidence, not a certificate of decay. Calling continuous monitoring "terminal" claims a closure that the logic itself denies.

Right, and this is the sharper point. Termination is certifiable by observation; non-termination is not. A signal's failure is certifiable the moment it fails badly enough to notice. Its continued health is never certifiable — only continually un-falsified. That asymmetry is exactly the shape of the halting problem, and continuous intake is the best observational posture available given that asymmetry, not a way around it. It does not tell a systematic PM "this signal is fine forever." It tells them, with provenance, exactly how long the current belief has gone un-falsified, on which feeds, since which regime shift, and lets that belief be revised the instant contrary evidence arrives rather than after a quarterly review. That is a semi-decision procedure, not a certificate. It is also, given the theorem, the only kind on offer.

What the third rung buys, concretely

The practical translation is a portfolio of live beliefs about each signal's decay state, each carrying where it came from and when it was last checked, rather than a single fitted model re-run periodically. That does not eliminate the systematic PM's judgement call about turning a strategy off. It changes what evidence exists at the moment the call has to be made, and whether the delay between decay and detection is measured in feed latency or in the length of a backtest window.

The budget objection, and where it actually lands

Unbounded streams and full provenance are not free. Reconciling every cross-asset feed against every prior belief costs storage, bandwidth and compute that scale with the number of streams, and in practice any real system forgets, compresses and samples — reintroducing a cutoff, just a rolling one. This is true and should not be minimised. The distinction that survives it is whether the cutoff is a design invariant, fixed and unrevisitable, or an operating parameter that can be widened, re-derived from retained provenance, or pointed back at history when a new question demands it. A model frozen at a training cutoff cannot be asked about a decay pattern that started the day after training ended, at any price. A live-intake system with a rolling window can be. That converts an absolute wall into a budget line. Budgets are hard, and every trading desk lives inside one. They are not the same kind of obstacle as a theorem, and the difference between the two is the entire content of the claim that this axis has a top rung.