A price-on-flow regression can overstate impact by up to 68% when hidden information moves both. Without that channel, it understates impact by 43%. Thumm and Chen prove the point inside a market they built. Their real-market evidence supports a narrower claim, as they acknowledge: it is "consistent with, though not proof of, causal transfer".

Where the impact estimate comes from

The target is the realized price path minus the path that would have occurred without the trade. This is Bonart's identity; we have examined its decay claims before (/articles/diffusive-market-impact-a-consistent-microfoundation). Almgren-Chriss and propagator fits instead estimate mean price conditional on flow. A common driver of flow and price separates that estimate from the causal effect.

The authors put such a driver into a three-variable stochastic system. Hidden "core" flow U pushes observable flow A and moves price Y directly, with coupling γ. Mechanical impact λ is the quantity a trader needs. Each episode has an exact dial for confounding strength, and the naive OLS slope has a closed-form bias linear in γ. Its measured γ-gradient is 0.326, close to the analytic 0.332. Even with γ at zero, the naive slope runs about 43% below true impact because price low-pass filters fluctuating flow. Raise γ to 2 and the error becomes a 68% overestimate. Hidden information determines which way the classical estimate misses.

Their prior-data fitted network, or PFN, is a transformer trained to predict post-intervention price quantiles from pre-onset context and a trade description. It has 4.2M parameters and takes about two hours to train on one A100. The training episodes are simulated, with exact shared-noise counterfactuals. U remains hidden, so data alone do not identify impact; the prior supplies identification. The authors make that dependence explicit. Their "do-ablated twin" is the same network with its five intervention inputs zeroed. Subtracting the twin's output from the causal output on identical episodes gives the "paired do-head effect", their main transfer readout.

The synthetic grid favors the PFN. Across γ, its bias ranges from −0.23 to +0.16; the oracle stays within 0.013. For γ ≥ 1 it has the best bias and RMSE among feasible methods. At γ=2, bias is +0.16, versus +0.32 for the square-root law and +0.26 for the propagator. The twin performs like the pre-window mean. Longer training barely helps: 4x the time moves γ=2 bias from 0.396 to 0.393. Shifting the training prior toward high confounding cuts it 57%, to 0.169.

What does ABIDES establish?

ABIDES supplies an actual counterfactual: a roughly 25 to 40% participation metaorder runs for 15 minutes, then the identical seed is replayed without it. Across 1,000 held-out seeds, measured impact is +0.28σ [0.05, 0.50]. The PFN gives +0.379σ [0.342, 0.416], with bias +0.08σ and an interval spanning zero. Classical estimates reach much the same level. The square-root law gives +0.42 and bias +0.13; the propagator gives +0.38 with a band about 4x wider. Here the PFN's advantage is interval width.

The authors arrived at unit standardization and soft encoding while iterating on seeds 0 to 999. They report the headline result from one frozen run on seeds 1000 to 1999. Encoding the metaorder as a clamped level rather than added flow drops the response to +0.10σ across 300 exploratory episodes, with bias still −0.31σ. They describe the roughly 4x ratio as indicative, rather than a matched ablation.

Averages are doing the work. Seed-replay noise has an SD of 3.6σ, leaving per-episode accuracy out of reach. Correlations with order size (0.09) and realized impact (0.11) reach the ceiling that noise permits. The response gets the average right while remaining, in the authors' words, strongly bimodal.

Real bursts leave causality open

The Binance perpetuals sample covers 91 days, April to June 2026. One-sided bursts are detected at |z| ≥ 5 relative to a trailing 90-minute window, yielding 763 BTC, 758 ETH and 757 SOL events. The do-head effects are +1.42, +1.53 and +1.22σ. A follow-up covering 3,802 Nasdaq bursts across 50 names finds +1.14σ [1.09, 1.19].

Those comparisons use factual outcomes. The transformer's signed biases are −0.84, −1.05 and −0.76, an improvement on last-value prediction at −1.05, −1.17 and −0.98. The zero-parameter square-root law does better: −0.41, −0.47 and −0.38. It also has the lowest raw RMSE, at 2.1 to 2.5σ. The authors call this "a bar the zero-shot PFN does not yet clear on real data."

Controls show that event labels drive the model's response. Randomly flipping signs on the Nasdaq bursts moves the measured directional return from +1.05 to −0.13σ; the causal model follows the false labels and creates error. At non-event BTC anchors, a fake intervention produces +1.35σ of model bias while the twin stays at −0.03. At real events, that response reduces bias from −2.29 to −0.95. Momentum continuation after a burst remains a possible explanation, which the authors acknowledge.

The recurrent variant closes part of the factual gap. It has 1.94M parameters and feeds the intervention in at every active step. A three-seed ensemble brings burst bias to +0.02, −0.22 and +0.13σ, compared with −0.41, −0.47 and −0.38 for the square-root law. ETH's interval has an upper bound of −0.00, so it does not span zero. Individual seeds range from −1.2 to +0.6σ and reach the measured scale on only two of three. Correlation with a 12-cell reconstructed metaorder benchmark rises from +0.68 to between +0.92 and +0.96, then holds at +0.97 on 40 held-out days. Deconfounding worsens: on-prior bias reaches +0.52 at γ=2, versus the transformer's +0.16. The authors attribute the trade-off to the training recipe and leave a model that can do both as an open problem.

The execution gap

Transfer results are in pre-onset volatility units on a queried bar. We did not find a conversion to basis points of realized execution cost. The recurrent model also needs a scale constant k for each domain: 0.49 to 0.53 in crypto and 0.049 to 0.096 in ABIDES. Crypto k is fitted to observational impact and can therefore carry that target's confounding. Transfer intervals fare poorly. The transformer's nominal 80% band covers 0.32 on ABIDES and 0.47 to 0.51 on crypto; the recurrent variant reaches 0.49 on ABIDES. Split-conformal recalibration raises coverage to 0.82 at about 4x the width. On real data, its guarantee applies only to factual outcomes.

Prior design is a further deployment concern. Adding jumps to the hidden core fits that revised prior well but shrinks the zero-shot response 3 to 8x.

We could not test this ourselves. The intervention adds signed flow, requiring trade prints with aggressor flags or metaorder records. Minute OHLCV bars have neither, and we hold no perpetuals data.

The synthetic deconfounding is real. So is the ABIDES result. A desk's own metaorders, priced in basis points with the causal model beating the square-root law, would change my view of the real-market claim.