Method
Fifty things happened in the market today. Which of them actually moved your position? That question is harder than prediction, because causes don't announce themselves and they don't stay put. The factor driving returns this month can be noise by next month, and a tool that reports relationships averaged over a decade tells you nothing about today.
So CommonQuant is built around a different question: what is causally live in the market right now, and how long will it stay that way?
Causal world modeling
We maintain a running dynamical model of the market in which flows, volatility structure, macro conditions, and narrative pressure are treated as candidate causes of returns. The causal status of each is re-estimated continuously. Think of it as a state-space model of the market as a physical system: observed effects on the surface, latent drivers underneath, and an inference engine connecting them.
The payoff is a world model that moves at the market's own speed. Causal roles reverse, influence decays, and a driver that mattered in March can be measurably dead by June — the model will say so, with a number.
Assimilative causal inference
The core technique comes from applied mathematics, published in Nature Communications (Andreou, Chen & Bollt, 2026). Assimilative causal inference was developed for turbulent physical systems, climate dynamics and extreme events, where causal structure shifts faster than classical methods can track it. Markets have the same character.
Classical causal tests run forward: does knowing X improve your forecast of Y? ACI inverts the problem. Using Bayesian data assimilation, it asks how much knowing the future path of returns would revise your estimate of a candidate driver's state right now. When future effects carry information about a present cause, that cause is genuinely coupled to the market rather than merely correlated with it. The framework runs a filter, which estimates using data up to now, against a smoother, which estimates using the full path. The divergence between them, measured in relative entropy, is the causal score at each instant.
Three properties make this the right tool for markets. It is instantaneous: you get a causal score per day, not an average over ten years. It works when the candidate cause is unobservable, which most real market drivers are. And it produces a causal influence range, a mathematically derived horizon for how long each driver's influence persists, so the model knows the difference between a force that shapes the next quarter and one that dies in a week.
We implemented the full framework natively in Rust and run it per-feature across the whole market: the conditional Gaussian state-space model, the filter, the smoother, the entropy metrics, and the influence-range computation.
The inference stack
Re-estimating the causal state of an entire market, hundreds of symbols and dozens of candidate drivers every day, is a computational problem before it is a mathematical one. Off-the-shelf data science stacks are too slow to do it interactively, so we built our own: a Rust engine for feature computation, filtering, regime detection, options pricing, and backtesting, with zero-copy data paths into the reasoning layer.
The numbers matter because they change what's possible. A full market pipeline runs in under a second cold and effectively instantly warm. Per-symbol feature computation takes milliseconds. An options price takes about fifty nanoseconds. None of this is for high-frequency trading, which we don't do. It's so that deep analysis is cheap enough to rerun on every question, every day, instead of being a batch job you wait for.
On top of the causal engine, the stack runs Bayesian online changepoint detection for regime breaks, transfer entropy and information-theoretic dependence measures, and a statistics layer with bootstrap inference, hidden Markov regime models, and dynamic correlation, all in the same native core. A causal link established in one regime is never assumed to survive into the next.
Validation
Every result that reaches you has been attacked first. Tests run walk-forward: fit on the past, judge on the future, never peek. Placebo tests check whether random inputs reproduce a finding. If they do, the finding was never real.
The strangest thing validation has taught us: in our swing research, most of the profit came from the exit rules, not the entry signals. We spent months on entries. We did not expect that either.
The graveyard
The platform is shaped as much by the ideas that died as by the ones that survived. These were design beliefs we held, tested, and killed, including some we wanted to be true — and each one left a mark you can still see in the product.
Verdicts should be graded, not binary
FALSIFIED
How strongly a signal fired predicted nothing beyond whether it fired at all — the correlation between conviction magnitude and outcomes was indistinguishable from zero. Every verdict in the platform is binary against pre-registered conditions because of this result.
News flow should gate decisions
REJECTED
Across more than a hundred thousand historical trades, filtering decisions through news flow removed good trades as often as bad ones. Narrative stayed in the platform as evidence to weigh, never as a veto.
The platform should work intraday
REJECTED
The measured edge was smaller than the friction at any realistic cost; whatever exists at that speed belongs to firms with co-located servers. The whole stack is built for swing horizons instead, where a thesis has time to resolve.
Nothing here guarantees a correct answer. It guarantees that when we tell you a thesis is supported, the claim has survived every way we know to kill it. And when we don't know, you'll see “insufficient evidence,” not a confident guess.
MarketFabric
MarketFabric is our first product: the systematic research and trading engine where this method meets real market data and real execution. It runs the full stack against a multi-hundred-symbol US equity universe. Every order passes a nine-check risk gate. Every decision is written to an append-only audit ledger. It has been running live paper trading since March, and it is the proving ground for everything on this page. No technique ships into CommonQuant until it has survived contact with the market inside MarketFabric.