The demarcation problem in algorithmic trading

The strongest objection first

A systematic portfolio manager who has read some philosophy of science will find this argument easy to dismantle, and they should get the chance to try. Popper's falsifiability criterion is not a live standard in philosophy of science; it was gutted by the Duhem–Quine thesis and finished off by Lakatos, who showed that working research programmes survive contact with anomalies by adjusting auxiliary assumptions, not by capitulating to the first bad print. No serious philosopher of science today sorts disciplines by asking whether a theory forbids an observation. Importing that criterion to rank machine architectures — Large Language Model against Large World Model against Large Universe Model — looks like reviving a corpse to win an argument about market data feeds.

And there's a second objection sitting right behind the first, more damaging because it is more concrete. Continuous intake is not obviously a virtue in trading. A book that ingests every order-book tick, every headline, every 8-K amendment and every cross-asset correlation shift in real time is a book exposed to sensor noise, spoofed quotes, adversarial order flow and correlated data errors across venues. A quant desk running on a stable, curated, backtested signal set — refreshed on a disciplined quarterly cycle rather than reacting to every tick — may in practice produce better-calibrated risk than one that revises continuously and ends up chasing artefacts. The firehose is not free lunch. It is a known way to overfit live capital to noise.

Both objections are real. Neither survives complete.

What the Popper objection gets right

The historical verdict is correct: strong falsificationism is dead as a demarcation criterion. Duhem and Quine showed that no single observation refutes a hypothesis on its own — a bad fill on a pairs trade could mean the spread model is wrong, or the borrow cost changed, or the venue had a data glitch, or the hedge ratio needs recalibration. Lakatos's picture of research programmes protecting a hard core with adjustable auxiliaries describes exactly how a trading desk actually behaves when a strategy underperforms: nobody drops the strategy on the first losing week, and shouldn't.

But that whole debate concerns how blame is assigned once a disconfirming observation has arrived. It assumes there is a new data point to attribute fault to. A frozen model has a problem one level prior to attribution: no new data point arrives at all. Consider a signal built from a static training window of order-book and filing data, deployed without further intake. If that signal's edge decays — say a merger-arbitrage spread that used to close in four days now takes eleven because the regulatory environment shifted — the frozen system has no channel by which that fact could reach it. There is nothing to argue about, no auxiliary to adjust, because there is no observation in the loop at all. The argument here needs only the thin, procedural residue that outlived Popper's collapse: an empirical claim about markets needs some route by which contrary evidence can still arrive. It does not need decisive, one-shot falsification. It needs a live wire.

What the noise objection gets right

The second objection is the one systematic PMs live with daily, and it deserves more than a rebuttal — it deserves the concession that it identifies the actual failure mode of naive continuous intake. A book that re-estimates every parameter on every tick will indeed chase noise: bid-ask bounce, a fat-fingered print, a headline algorithm's false positive on a news feed, a data vendor's timestamp error propagating through a dozen downstream models. "Continuous" is not automatically "correct." Plenty of blown-up books were continuously updating right up until the blow-up.

The feature that separates useful continuous intake from thrashing is not volume of data but provenance and calibration attached to each stream. A well-built signal-decay monitor does not treat every tick as equally informative; it tracks, per signal, the origin of each input (which venue, which feed, which filing type), a rolling estimate of that signal's live hit rate against its backtested hit rate, and a confidence band that narrows or widens as evidence accrues. When a stream turns unreliable — a news-sentiment feed starts scoring headlines inconsistently after a vendor changes its NLP pipeline — that stream can be down-weighted or cut without touching the rest of the book, and every position derived from it can be re-priced with the bad input excised. That is expensive infrastructure, built and maintained badly more often than not. But it is a different claim from "ingest everything and average it," and the objection is right to reject the latter while leaving the former standing.

The characteristic failure this explains

Here is where the domain does the philosophical work rather than illustrating it after the fact. The signature failure of systematic trading is not a single bad trade. It is a signal that decays silently and keeps trading after it has stopped working. A statistical arbitrage pair that relied on a stable cointegrating relationship between two names starts drifting apart for structural reasons — one company's business mix has shifted, one index has rebalanced them out of the same sector — and the model, trained once on the historical relationship, keeps sizing positions as if the relationship still holds. Losses accrue not as a shock but as a slow bleed: a few basis points a day, absorbed by risk limits until they aren't. By the time monthly attribution flags the strategy as a persistent drag rather than a bad-luck streak, the systematic PM responsible has been running a model that stopped being an empirical claim about the market weeks earlier and became, without anyone deciding so, a recital of a pattern that used to be true.

That is the frozen-corpus failure mode transplanted into basis points. The model's assertion — "this spread mean-reverts" — became unfalsifiable in the operational sense the moment intake stopped tracking the specific regime that made it true, even though the backtest behind it was rigorous and the original fit excellent. Nothing about the model was false when built. The defect was structural: no channel existed for the world's drift to reach the position sizing before the P&L did.

The three regimes, in trading terms

A quarterly-refit factor model sits closer to the first row than desks like to admit: refitting on a calendar cadence is a scene that opens once a quarter and closes for the ninety days in between. A real-time signal-health dashboard that tracks each input's live decay against backtest and automatically retires stale signals is an attempt at the third. Most production trading systems are somewhere in between, and the honest self-description for most of them is closer to the second row than the third.

The narrower claim that survives

Continuous intake with provenance does not make a trading model correct, well-calibrated, or safe from noise. Plenty of continuously updating books have failed for exactly the reasons the second objection names. What continuous intake with provenance changes is whether a claim like "this spread mean-reverts" remains, in principle, breakable by next week's filing or next month's cross-asset correlation shift — and whether, when it breaks, the desk can trace the break back to the specific stream that carried the bad news, rather than discovering the failure only in the P&L.

Frozen models remain the right tool for the settled: microstructure facts about a specific historical episode, tick-size regime effects that are not going to reoccur, calibration exercises against closed datasets where nothing is expected to change. The claim is confined to signals whose truth-value moves with the market — which is most of them, held responsible by a systematic PM who has no other honest description of the job than staying answerable to evidence that has not arrived yet.