Market Impact Estimation via Causal Foundation Models

arXiv:2610.12537 · cs.LG, cs.CE, q-fin.TR · Submitted 2026-10-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Market Impact Estimation via Causal Foundation Models".

Jane: The gist The continuous-time causal prior-data fitted network (PFN) provides a proof-of-concept for estimating market impact by modeling hidden informational flow,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, we're wrapping up on "Market Impact Estimation via Causal Foundation Models." Basically, this paper is about using these new AI models to figure out how much a big trade actually messes with the price of an asset, and they’re using cause-and-effect math to do it.

Jane: Right. So, what does that mean for us? It means instead of just guessing based on old formulas, we can try to model the hidden stuff—like information flowing between traders—that classical models always miss.

Lu: The authors use this continuous-time causal prior-data fitted network, or PFN, trained on a synthetic market where they build in that hidden informational flow by design.

Meng: So they are basically building a computer simulation of a market with built-in confusion to see if the AI can untangle it.

Lalam: And on that synthetic setup, the PFN actually manages to track the true underlying parameters when things get really messy, showing a closed form for the error in their naive estimation.

Tom: That's neat because it shows a clear path for how this framework should work under high confounding conditions. But what about real markets?

Jane: They show zero-shot transfer, which means you don't need to retrain the model on new data just to make it work in a live environment, even responding to actual market orders.

Lu: They achieved that transfer by standardizing the scale of the input and making sure they treat those inputs as additive flow rather than just a fixed price level.

Meng: That’s what I care about—if an AI can jump into production without weeks of retraining, that's how we get value from these things. But what about its limitations?

Lalam: They did find that the recurrent variant, which tries to model the impact over time, still has some shape error left; it’s good at getting the overall size right but maybe not the exact pattern.

Tom: So even with these advances, we're still looking at a level factor error instead of a perfect shape capture. It shows that while causality is powerful, modeling every tiny detail of execution impact is still hard for these models right now.

Jane: Exactly. The big implication is that this causal approach gives us a much more honest measure of what it really costs to trade, moving us closer to understanding market mechanics beyond just simple correlation.

Lu: It identifies identification through the prior structure, which is a strong way to supply that necessary link between cause and effect when you don't have perfect data.

Meng: It’s a proof-of-concept for how we can build more robust execution strategies, even if it still needs refinement on the shape estimation side before we can ship it widely.

Conclusion: Tom: So we're wrapping up on "Market Impact Estimation via Causal Foundation Models." Basically, this paper is about using these new AI models to figure out how much a big trade actually messes with the price of an asset, and they’re using cause-and-effect math to do it.

Jane: Right. So what does that mean for us? It means instead of just guessing based on old formulas, we can try to model the hidden stuff—like information flowing between traders—that classical models always miss.

Lu: The authors use this continuous-time causal prior-data fitted network, or PFN, trained on a synthetic market where there’s a hidden informational flow that confuses the relationship between the trading flow and the actual price movement.

Meng: So they're building this simulator with an exact knob for how strong that confusion is per episode, which is interesting because it lets them see if their model can handle a lot of confounding.

Lalam: And on this synthetic setup, they show that the naive estimator’s bias follows a closed form, and the PFN actually manages to track the unbiased true parameter oracle when the confounding strength is high.

Tom: That's a big win for accuracy on this simulated data—they confirmed that up to sixty-eight percent overestimation, which is pretty solid evidence of its potential.

Jane: But what makes this paper really compelling for us here on the show is how they test if this causal understanding actually transfers to real markets, and that’s where things get really interesting with the zero-shot results.

Lu: They showed zero-shot transfer, meaning you don't even retrain the model at all; it can respond at the ground truth scale to metaorders in a limit order book, like an ABIDES book.

Meng: Zero-shot transfer is tough to prove because real markets are messy and unpredictable, so if it works out of the box without retraining, that’s a strong indicator.

Tom: The paper claims it responds at the ground truth scale with a plus zero point three eight standard deviations against seed-replay counterfactuals in that limit order book. That number is what we need to look at for real performance compared to existing methods.

Wrap-up: Jane: And they also showed it reduces prediction bias around detected flow bursts on BTC, ETH, and SOL perpetuals by a paired do-head effect of plus one point two to one point five standard deviations over about seven hundred sixty events each in Table one.

Lu: They found two specific mechanisms that enable this zero-shot transfer: first, prior-scale standardization, and second, matching the intervention's semantics—meaning they treat the input as an additive flow rather than just a clamp on price levels.

Meng: Matching the semantics is key because in real trading you’re usually adding more volume or flow to a market rather than just setting a fixed price barrier.

Tom: And then they introduced this recurrent state-space variant, which is where it gets even more complex, and that’s what really closes the shape gap in the impact estimation.

Jane: That recurrent variant uses stacked gated delta-product recurrences, and they train it with a fifty/fifty soft-intervention mix to build up the impact over execution duration.

Lu: Because in a transformer architecture, the response is usually one shot, so this recurrent structure allows the model to compose the impact in its state across time steps.

Meng: So instead of predicting one number for a trade, it’s predicting how that trade compounds as it runs through the book. That makes sense for real execution dynamics.

Tom: And on this variant, they saw the correlation with measured metaorder impact go from plus zero point six eight down to between plus zero point nine two and plus zero point nine six across training episodes, which is a significant lift in accuracy over the original transformer approach.

Jane: That’s a big jump in correlation, moving it much closer to what we measure when we look at actual market impact across different assets like BTC/ETH/SOL perpetuals and Nasdaq equities.

Lu: They also found that on real markets, specifically with this causal M-DoT-PFN, the paired do-head effect is positive across all assets, which is consistent with what we expect to see when you compare the actual execution path to a counterfactual path.

Wrap-up: Meng: It’s good they're looking at real market performance on BTC/ETH/SOL and Nasdaq equities because that’s where we actually deploy these models, not just in simulation.

Tom: However, they also found that the zero-shot square-root law sits closer to zero on all three assets, with values of minus zero point four one for BTC, minus zero point four seven for ETH, and minus zero point three eight for SOL when looking at the robust range of the literature prefactor.

Jane: So while it’s better than Almgren–Chriss or propagator models because it’s causal, these zero-shot results still show some noise, with those negative values being relatively close to zero.

Lu: And they also noted that the recurrent variant's remaining error is a level factor rather than a shape factor, meaning it’s not necessarily capturing the precise path of impact shape perfectly yet.

Meng: That tells me we might still need more work on refining how it captures the exact *shape* of the execution impact, even if it gets closer to tracking the overall magnitude better.

Tom: The paper concludes that the consistent through-line is this positive paired do-head effect across synthetic, agent-based, and real data when using this causal inference under hidden confounding via a prior.

Jane: So what does that mean for someone just listening to the show? It means we have a new way to estimate market impact that tries to account for hidden information channels, which is something classical models totally miss by only looking at the flow and price.

Lu: The implication is that if this framework works on real data, it provides a more robust estimate of what it costs to move large amounts of assets in a live environment than the older regression-based methods.

Meng: For deployment right now, we’d need to make sure the recurrent variant doesn't require too much compute at scale, especially since they trained that recurrent state-space model with one point nine four million parameters.

Tom: So yeah, it’s a powerful proof-of-concept that uses causal modeling to bridge the gap between what we think is happening and what actually happens when you execute a large order in a live market environment.

Dennis Thumm, Ying Chen

National University of Singapore

cs.LG, cs.CE, q-fin.TR

Submitted: 2026-10-08

Updated: 2026-10-08

Comments: In 7th ACM International Conference on AI in Finance (ICAIF '26), November 14-17, 2026, Milan, Italy

Code: https://github.com/thummd/causal-marketsimulator

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 86/100

The gist: The gist The continuous-time causal prior-data fitted network (PFN) provides a proof-of-concept for estimating market impact by modeling hidden informational flow, which allows it to track the

Key concepts

Market Impact: This is the difference between the actual price a large trade causes and what the price would have been if the trade never happened. Classical models struggle because they assume a direct relationship between trading volume and price change, often absorbing information channels that introduce bias.
Causal Prior Structure: The model is built on a synthetic market prior where hidden informational flow is intentionally designed to confound the relationship between observable trading flow and price movement. This structure allows the model to learn how information moves through the market alongside trades.
PFN (Prior-Data Fitted Network): This is a continuous-time network trained on both synthetic market data and real data. It uses a transformer architecture to encode observations and an intervention mechanism to simulate 'do' operations, allowing it to estimate impact even when the relationship is complicated by hidden factors.

Terminology

Foundation Model — 1× in this paper · explained in 3 episodes
Foundation Models — 1× in this paper · explained in 7 episodes

Summary

The gist The continuous-time causal prior-data fitted network (PFN) provides a proof-of-concept for estimating market impact by modeling hidden informational flow, which allows it to track the unbiased true parameter oracle at high confounding levels and transfer its causal capability zero-shot into real markets.

Market Impact Definition and Classical Models

Market impact is defined as the realized price path minus the counterfactual path that would have occurred without the trade. Classical impact models like Almgren–Chriss and propagator models fit linear permanent-impact coefficients by regressing price on flow or fitting a decay kernel to observational data. Estimated this way, impact coefficients absorb the information channel, and execution-cost models inherit the bias.

The Causal Prior Structure

The paper introduces a market prior where hidden informational flow U confounds the flow–price relation by construction. The flow U drives observable flow A and moves the price Y through the channel γ. The confounding strength is parameterized by a core share s. The naive estimator’s bias in closed form is given by biasOLS(γ) = γ Cov(A,U) + w Var(U)/(thetaY + thetaU).

Model Architecture and Causal Inference

The continuous-time causal PFN is built upon a transformer architecture that encodes pre-intervention observations and masks post-onset observations. The intervention specification (target, kind, value, window) enters a cross-variable mixer as a distinguished channel. The model amortizes the posterior predictive interventional distribution p(Y (tq) do(·), context), marginalizing over the latent Structural Causal Model (SCM) given the observed context. A recurrent state-space variant is used to compose impact in its recurrent state, where its training recipe adds to the prior of §4 a 50/50 soft-intervention mix.

Transfer Capabilities and Robustness

The paper demonstrates zero-shot transfer by showing the model responds at the groundtruth scale to metaorders in an ABIDES limit-order book. Two mechanism findings enable this transfer: prior-scale standardization and matching the intervention’s semantics (additive flow, not a clamp) to the event’s. The recurrent variant's response accrues over execution duration where the transformer’s stays flat, raising the correlation with measured metaorder impact from +0.68 to between +0.92 and +0.96.

Real Market Performance and Limitations

On real markets, specifically BTC/ETH/SOL perpetuals and Nasdaq equities, the paired do-head effect is observed to be positive across all assets. The causal M-DoT-PFN is less biased than Almgren–Chriss, the fitted propagator, and last-value prediction on every asset. However, the zero-shot square-root law sits closer to zero on all three (−0.41, −0.47, −0.38; robust to the literature prefactor range). The recurrent variant's remaining error is a level factor, not a shape factor. The paper concludes that the consistent through-line of a positive paired do-head effect on synthetic, agent-based, and real data while the ablated twin collapses onto the mean supports the narrower, load-bearing claim. The framework is load-bearing because it identifies causal inference under hidden confounding by supplying identification via the prior <ref:2610.

Improvements for AI systems

Bold Header: Causal Market Simulation via Prior-Data Fitted Networks (PFN)

The improved system can estimate market impact by learning what would the price have been had I not traded using a continuous-time causal prior-data fitted network trained on a synthetic market prior where a hidden informational flow confounds the flow–price relation by construction.

Bold Header: Implicit Deconfounding and Oracle Tracking

The system will track the unbiased true-parameter oracle at high confounding, as the text states, while classical models suffer from bias because they do not account for flows that move prices through information as well as trading.

Bold Header: Zero-Shot Transfer to Real Market Data

The model can respond at the groundtruth scale to metaorders in an ABIDES limit-order book and replicate on a large number of detected flow bursts in BTC/ETH/SOL perpetuals, demonstrating that Transfer hinges on encoding the metaorder as added flow rather than a clamped price level.

Bold Header: Duration-Dependent Impact Modeling via Recurrent State-Space Variant

A recurrent state-space variant will be used to capture how impact accrues over execution duration where the transformer’s stays flat, raising correlations with measured metaorder impact from +0.68 to between +0.92 and +0.96.

Bold Header: Causal Calibration of Counterfactual Uncertainty

The system can utilize split-conformal recalibration to address uncertainty, as the paper notes that the interval calibration does not survive transfer (§6.3) even where point predictions do, specifically restoring nominal-80% coverage on counterfactual ground truth in ABIDES.

Bold Header: Real Market Impact Estimation from Detected Flow Bursts

The system can predict the realized postevent price given the event’s signed size on real market data (BTC/ETH/SOL) by using detected flow bursts as interventions, achieving a paired do-head effect of +1.42 sigmaˆ (BTC).

Bold Header: Distinguishing Causal vs. Predictive Bias

The system will use the do-pathway, not the sequence model, carries the causal work, ensuring that its estimates are based on what happens if I trade rather than merely predicting what typically happens next.

Abstract

When a trader executes a large order, the relevant question is not what will the price be? but what would the price have been had I not traded? Market impact is thus a causal quantity, yet classical models (Almgren--Chriss, propagators) regress price on flow, which is biased whenever informed flow moves prices through information as well as trading. We present a proof-of-concept Causal Market Simulator: a continuous-time causal prior-data fitted network (PFN) trained on a synthetic market prior in which a hidden informational flow confounds the flow--price relation by construction, with confounding strength an exact per-episode knob. On this prior, the naive estimator's bias follows a closed form we confirm empirically (up to 68% overestimation), while the PFN tracks the unbiased true-parameter oracle at high confounding. Its do-ablated twin, the same model with the intervention inputs removed, collapses onto the pre-window mean. Zero-shot, the model responds at the ground-truth scale to metaorders in an ABIDES limit-order book (+0.38 pre-onset standard deviations against seed-replay counterfactuals), and its intervention response replicates on about 2,300 detected flow bursts in BTC/ETH/SOL perpetuals and 3,800 on Nasdaq equities, evidence consistent with, though not proof of, causal transfer. Off-prior the response degrades gracefully, and split-conformal recalibration repairs interval calibration. Transfer hinges on encoding the metaorder as added flow rather than a clamped price level, not on more training. A recurrent state-space variant closes the remaining shape gap: its impact accrues over execution duration where the transformer's stays flat, raising the correlation with measured metaorder impact from +0.68 to between +0.92 and +0.96, at the cost of weaker on-prior deconfounding that traces to the training recipe rather than to recurrence.