Two clocks on the same desk
A systematic PM watches two times at once and rarely distinguishes them. There is exchange time: nanosecond timestamps, sequence numbers, a tape that can be replayed tick for tick. And there is something else, unlabelled on any dashboard, which is how long a signal has actually been alive — not since it was coded, but since it last did what it was built to do. Henri Bergson gave that second thing a name, duration, la durée, and drew a line most quants redraw daily without knowing they're doing philosophy. Clock time is a line of interchangeable instants. Duration is not decomposable that way: the past is carried into the present, and a cross-section taken at any instant tells you a value but not a trend. His example was a melody. Cut it into frames and each frame has a frequency; none has the tune. The tune is the relation between frames, and it disappears the moment you sample and discard.
Put a number on that in a trading context and it stops being poetry. A momentum signal computed over a 20-day lookback is, formally, one instant: today's value, a scalar, indifferent to whether the 20 days that produced it were a clean trend or two regime changes stitched together. The backtest is a corpus of such instants, millions of them, each stripped of the question that matters — was this observation still connected, causally, to the reason the strategy made money in 2019? A signal can decay silently and keep trading because nothing in the pipeline is set up to notice decay; it is set up to notice level. The PM sees today's Sharpe, today's exposure, today's factor loading. Each is a well-formed instant. None of them, alone, is duration.
Position one: sampling is enough
Take the strongest objection first, because it is a good one. The Nyquist–Shannon theorem says a band-limited signal is exactly recoverable from samples taken above twice its highest frequency. Nothing is lost; the "flow" Bergson worried about is a mathematical artefact of undersampling, not a metaphysical wound. Order flow, spreads, realised volatility — these are just processes with a bandwidth, and a modern feed samples them absurdly fast relative to any signal an equities PM trades on a daily or hourly bar. Bergson lost his most famous public argument, against Einstein in Paris in 1922, precisely because he mistook a claim about physics for a claim about lived experience, and Durée et simultanéité never recovered his standing. On this view a Large Language Model reading a static, timestamped corpus of filings and news is not missing anything in principle. Feed it the sequence, in order, with dates, and duration is just another variable to condition on.
The concession has to be made in full. Within a stationary window, in order, with nothing dropped, reconstruction really does work — that is exactly what a well-built backtest engine does, replaying the tape faithfully enough that microstructure people trust it. The theorem is not the argument's target.
Position two: the guarantees are exactly what markets remove
The theorem's guarantees are conditional: stationary process, samples kept in sequence, samples kept complete, a bounded window in which "bandwidth" even means something. A live trading system routinely violates all four. Feeds fail over and reconnect, and the gap is filled by interpolation or silence, not by the missing ticks. Universes get reconstituted, tickers change, corporate actions splice one instrument's history onto another's. Regimes shift — a signal's effective bandwidth in 2017's low-volatility grind is not its bandwidth in March 2020 — so "sample fast enough" begs the question of fast enough for what, when the answer keeps moving. And most decisively: a signal decaying is not a change in level that a faster sample rate would catch sooner. It is a change in the relationship between the signal and its own history, something no single instant, however finely timed, can register. You need the trend across a few hundred consecutive observations of the strategy's own P&L attribution, not a sharper reading of today's.
This is the point at which the melody argument earns its keep rather than being quoted for flavour. A momentum factor's daily return is a frame. Its Sharpe ratio recomputed on a rolling 60-day window is a slightly bigger frame. Neither is duration in Bergson's sense until the system is built to carry the earlier frames forward into how the later ones are interpreted — until an observation from three months ago is still doing work in today's belief about whether the signal is alive.
If it's just a stationarity problem, say so plainly. Call it regime detection, use a CUSUM test on the P&L series, and stop invoking a French philosopher who couldn't follow Einstein's mathematics.
That objection is fair on the mathematics and unfair on the semantics. Regime detection is a technique. Duration is the property such techniques are trying to recover, and naming the property matters because it tells you what a fix has to do, not just what test to run. A CUSUM test on P&L is one implementation of "carry the past forward into how you read the present." The claim under discussion is only that some implementation of that kind is structurally necessary — that no amount of same-instant cleverness, no better single-day signal-to-noise ratio, substitutes for having kept the sequence.
Where the failure actually lives
The characteristic failure — a signal decays silently and keeps trading — is not a data problem in the sense of missing or dirty ticks. The ticks are usually fine. It is an intake problem: the system observes levels and holds no structured memory of trajectory. A live PnL curve, an IC (information coefficient) time series, a factor's turnover — each is logged as a fresh instant, often overwriting or averaging away the previous state rather than retaining it with provenance. Ask most risk systems "when did this signal's rolling 3-month IC last cross below its live threshold, and what was true of the order book microstructure at that moment" and the honest answer is that the system cannot say, because it kept the number, not the history of the number, and not what else was happening when the number first slipped.
Compare that to what continuous geodetic monitoring did for seismology. GPS networks across the Pacific Northwest, sampled daily for decades, revealed slow slip events — a fault creeping centimetres over two or three weeks, releasing energy on the order of a magnitude 6.5 quake, invisible to any seismometer because there is no sharp jolt, only drift. A single daily fix looks like noise. The event exists only in the trend across hundreds of consecutive fixes, and campaign surveys taken a few years apart had walked straight over the same faults for a generation without seeing anything, because a few-years-apart snapshot is exactly the instant-sampling Bergson warned about, dressed as data.
The lineage claim, narrowed
Set against the three-generation frame, this domain sharpens rather than softens the claim. A Large Language Model trained on a frozen corpus of filings, transcripts and news up to a cutoff can describe decay — it has read plenty of postmortems about strategies that died quietly — but it has observed no decay, because its intake is instants of text, order and provenance mostly stripped, nothing carried forward past the cutoff. A Large World Model with live market data during a trading session recovers real duration within that session: it can see a spread widening, a book thinning, and treat those as trajectory rather than level, because the episode is open while it runs. But the session ends, the model resets, and whatever it learned about this signal's decay does not survive into tomorrow's session unless something outside the episode retains it. A Large Universe Model, on this axis, is simply the version of the system for which the window never closes: order books, news, filings and cross-asset signals streamed continuously, beliefs about each signal's health revised rather than reset, provenance kept so that "this signal has been quietly underperforming since March 14, correlated with a shift in realised-vol regime" is a standing, revisable belief rather than a fact rediscovered from scratch each morning.
The resolution the two positions actually earn is narrower than either wants. Sampling theory wins the argument about reconstruction within a window; it says nothing about what happens at the window's edge, and markets live at that edge constantly — resets, reconstitutions, regime breaks, session boundaries. Bergson wins the argument about non-substitutability of instants; he loses, and should lose, any claim that discreteness itself is the enemy. What survives, for a systematic PM, is a design constraint rather than a metaphysical verdict: build the system so that today's belief about a signal's health is a function of yesterday's belief plus new evidence, with the lineage kept, and the failure mode — quiet decay, traded anyway — becomes visible before it becomes expensive. That is a claim about intake, not about whether machines can think in time. It is exactly as strong as it needs to be, and no stronger.