Practitioner synthesis. A propagator kernel estimated from public trades and quotes is a correct answer to a narrow question: what linear filter of past order signs best reproduces returns. It is not, on that evidence alone, a measurement of how the impact of your order decays. When returns are unpredictable from past signs, which is roughly what an efficient market looks like, the shape of the estimated kernel is fixed by the autocorrelation of order signs and nothing else. A model in which impact never decays produces the same estimate. The worked numbers below are closed-form arithmetic on a stylized sign process. They illustrate the mechanism and are not calibrated to any venue.

The problem: a kernel that fits and a scheduler that misses

The workflow is standard. Take a year of public tape, sign the trades, estimate an impact-decay kernel, and hand it to a scheduler. In an Almgren–Chriss setup the kernel collapses to a permanent and an instantaneous coefficient. In an Obizhaeva–Wang setup it becomes a resilience rate. In the general transient-impact setup it enters the cost functional directly, and the optimal schedule solves a Fredholm equation in the kernel. The fit is good, since the response function is reproduced almost by construction. Then live implementation shortfall drifts away from the backtest: the predicted post-trade reversion does not arrive in full, and schedules that were supposed to be cheap because impact “relaxes between child orders” are not.

The usual suspects are a stale calibration window or a regime change. There is a more basic candidate. The kernel was never identified as a property of a single order’s impact. A recent paper in the corpus, Fujiwara, Sato and Kanazawa (2026), states this as a no-go theorem: under long-memory order flow and diffusive prices, propagator estimation returns a transient, fine-tuned decay whether or not the true impact decays. The rest of this essay works through why, in discrete time and with elementary tools, and what to do about it.

The linear propagator and its normal equations

Let \(\epsilon_t \in \{+1, -1\}\) be the sign of the \(t\)-th market order and \(\Delta P_t\) the mid-price change it is associated with. The discrete-time linear propagator, or transient impact model (TIM), is

$$ \Delta P_t = \sum_{\tau \ge 0} G(\tau)\,\epsilon_{t-\tau} + \eta_t . $$

Here \(G(\tau)\) is the kernel on returns. The quantity people draw when they say “impact decay” is the cumulative response of the price level to one isolated trade,

$$ \bar G(\ell) = \sum_{\tau = 0}^{\ell - 1} G(\tau), $$

so that \(\bar G(1) = G(0)\) is the immediate impact and \(\bar G(\infty)\) is whatever is permanent. Decay of \(\bar G\) means that \(G(\tau)\) is negative for \(\tau \ge 1\).

Two observable second moments carry the estimation:

$$ C(\ell) = \mathbb{E}[\epsilon_t\,\epsilon_{t-\ell}], \qquad \mathcal{R}(\ell) = \mathbb{E}[\Delta P_t\,\epsilon_{t-\ell}] . $$

Multiply the model by \(\epsilon_{t-\ell}\) and take expectations. If the residual is uncorrelated with current and past signs, \(\mathbb{E}[\eta_t\,\epsilon_{t-\ell}] = 0\) for \(\ell \ge 0\), then

$$ \mathcal{R}(\ell) = \sum_{\tau \ge 0} G(\tau)\,C(\ell - \tau), \qquad \ell \ge 0 . $$

These are Yule–Walker equations. On the half-line they form a discrete Wiener–Hopf problem. Truncated at \(L\) lags they are a Toeplitz system \(\mathbf{C}\,\mathbf{g} = \mathbf{r}\), and because an autocovariance matrix is positive definite for any non-degenerate sign process, the solution exists and is unique.

That uniqueness is where the trouble starts, because it is easy to read too much into it. The solution \(\mathbf{g}\) is the unique linear projection of returns on past signs. It equals the structural response of price to a trade only under two further conditions that the equations cannot check:

  1. Exogeneity. Signs are not chosen in response to the same information that moves \(\eta_t\). Any trader who buys because the price is about to rise violates this.
  2. History independence. The response to a trade does not depend on how predictable that trade was. Every trade, anticipated or not, receives the same \(G\), and \(G\) stays the same if the law of the order flow changes.

Under both conditions the deconvolution identifies the kernel. Drop the second and the same numbers admit a different reading.

The observational-equivalence trap

Consider the opposite structural story. Impact is permanent, but only the unexpected part of the order flow moves the price:

$$ \Delta P_t = \theta\,\big(\epsilon_t - \hat\epsilon_t\big) + \eta_t, \qquad \hat\epsilon_t = \sum_{k \ge 1} a_k\,\epsilon_{t-k}, $$

where \(\hat\epsilon_t\) is the best linear forecast of the next sign from past signs. This is the logic of Hasbrouck’s (1991) vector autoregression, where the information content of a trade is the long-run price response to the trade innovation. It is also the Lillo–Farmer picture of long-memory flow met by liquidity that adjusts to it, and what Taranto, Bormetti, Bouchaud, Lillo and Tóth (2018) formalize as the history-dependent impact model (HDIM). The mechanism is reactive liquidity: when a buy is expected, the ask side has already been replenished or repriced, so the expected trade arrives and nothing more happens.

Expand the forecast and compare with the propagator:

$$ \Delta P_t = \theta\,\epsilon_t - \theta \sum_{k \ge 1} a_k\,\epsilon_{t-k} + \eta_t \quad\Longrightarrow\quad G(0) = \theta, \qquad G(k) = -\theta\,a_k \quad (k \ge 1). $$

The permanent-impact model is a propagator model, with a kernel that decays because the forecast weights \(a_k\) are positive for persistent flow. The level response is

$$ \bar G(\ell) = \theta \Big( 1 - \sum_{k=1}^{\ell-1} a_k \Big). $$

On a single stream of signed trades the two models generate the same \(\mathcal{R}(\ell)\) and the same \(C(\ell)\). No second-moment statistic of the public tape separates them.

Diffusive prices make the kernel a function of the signs alone

The equivalence is not a coincidence between two hand-picked models. Suppose only that returns cannot be predicted from past order signs:

$$ \mathcal{R}(\ell) = 0 \qquad \text{for all } \ell \ge 1 . $$

This is the diffusivity condition, and it is close to what liquid markets deliver. Substitute it into the Yule–Walker equations for \(\ell \ge 1\) and divide by \(G(0)\):

$$ C(\ell) = \sum_{k \ge 1} \Big( -\frac{G(k)}{G(0)} \Big)\, C(\ell - k), \qquad \ell \ge 1 . $$

These are exactly the normal equations that define the linear forecast weights \(a_k\) of the sign process. So for any data-generating process with unpredictable returns,

$$ \frac{G(k)}{G(0)} = -a_k , $$

and the estimated kernel is the prediction-error filter of the order signs. Price data enter only through the scale: with \(\sigma_e^2 = C(0) - \sum_{k \ge 1} a_k C(k)\) the one-step forecast-error variance of the signs, the \(\ell = 0\) equation gives \(G(0) = \mathcal{R}(0) / \sigma_e^2\). Everything about the shape of the decay was already in \(C\) before a single price was read.

A worked case: long memory in, fine-tuned decay out

Take a sign autocorrelation with the long memory seen in equity order flow, \(C(\ell) \sim \ell^{-\alpha}\) with \(0 < \alpha < 1\), and for concreteness the fractional-noise form with \(d = (1 - \alpha)/2\). Its forecast weights are known in closed form, sum to one, and decay as \(a_k \sim k^{-1-d}\). The estimated level kernel is therefore

$$ \frac{\bar G(\ell)}{G(0)} = \sum_{k \ge \ell} a_k = \frac{\Gamma(\ell - d)}{\Gamma(\ell)\,\Gamma(1 - d)} \;\sim\; \frac{\ell^{-d}}{\Gamma(1-d)}, \qquad \gamma = d = \frac{1 - \alpha}{2}. $$

That is the Bouchaud–Gefen–Potters–Wyart fine-balance exponent, arriving with no assumption about impact decay at all. For \(\alpha = 0.5\), a typical order of magnitude for large-cap equities:

Trades after the event \(\ell\)125101001,000
Estimated \(\bar G(\ell)/G(0)\)1.000.750.560.470.260.15

The estimate says that 85% of a trade’s impact is gone after a thousand trades and that all of it goes eventually, since \(\sum_k a_k = 1\) forces \(\bar G(\infty) = 0\). The same table is produced when the truth is \(\theta\) of permanent impact per unit of surprise. The fitted scale is also off from the naive reading: for unit-variance signs with this autocorrelation \(\sigma_e^2 = \Gamma(1-d)^2 / \Gamma(1-2d) \approx 0.85\), so \(G(0) \approx 1.18\,\mathcal{R}(0)\).

This is the discrete, linear skeleton of the Fujiwara–Sato–Kanazawa result. Their theorems are stronger. They show that the Wiener–Hopf deconvolution under long-memory flow necessarily returns a transient propagator, and they exhibit an exactly solvable null model with order splitting and a square-root law in which impact is strictly permanent while the inferred propagator still shows the fine-tuned decay. They also note that nonlinear propagator estimation inherits the problem when metaorders themselves have long memory. Our score for the paper is high on math and low on empirical rigor, because the argument is a proof plus a null model, not a test on market data.

Where the two readings disagree

If both stories are the same linear filter, why does it matter which is true? Because an execution scheduler does not replay the historical flow. It injects a new flow with a different law, and the two stories extrapolate differently.

  • Under the transient reading the kernel is a property of the market. Your schedule \(x_1, \dots, x_T\) pays, on top of each child’s own immediate impact, an expected cost \(\sum_{s < t} x_t\,\bar G(t - s)\,x_s\) for trading into the residue of its earlier children. Spacing out child orders is rewarded because impact relaxes between them, and the price after completion reverts along \(\bar G\) toward zero.
  • Under the permanent reading the weights \(a_k\) describe how the market forecasts flow. What your next child order costs depends on how well the market has detected your parent order. A detected metaorder pays little for its later children. An undetected one pays \(\theta\) on each. After completion the price reverts only by the amount the market had over-anticipated your continuation.

These give different optimal schedules and different shortfall forecasts, and public tape cannot arbitrate, because the parent-order labels that distinguish “a persistent buyer” from “many unrelated buyers” are not on it. This is the Lucas critique in execution form: the estimated coefficients are reduced-form, and they are not guaranteed to survive a change in your own policy.

What no-dynamic-arbitrage rules out, and what it leaves open

A natural hope is that theory closes the gap: perhaps only one of the readings is free of price manipulation. It does not, and it is worth being exact about which condition does what, because two different results tend to get merged.

The efficiency condition. With \(C(\ell) \sim \ell^{-\alpha}\) and \(\bar G(\ell) \sim \ell^{-\gamma}\), the price variance grows linearly only if

$$ \gamma = \frac{1 - \alpha}{2} . $$

A slower decay gives superdiffusion, meaning trends that anyone watching the tape could trade. A faster one gives subdiffusion, meaning mean reversion. This is the Bouchaud–Gefen–Potters–Wyart fine balance. It is a statement about statistical arbitrage for observers of the flow, it holds as an equality and not a bound, and the previous section shows why it carries no evidential weight: the estimator returns this exponent whenever prices are diffusive.

Gatheral’s condition. Gatheral (2010) asks a different question: can a trader run a round trip with negative expected cost? In his continuous-time model, where trading at rate \(v\) moves the price by \(f(v)\) and the effect decays through a kernel, the answers include: exponential decay is consistent only with linear instantaneous impact, and for power-law impact \(f(v) \propto v^{\delta}\) with power-law decay \(\tau^{-\gamma}\),

$$ \gamma + \delta \ge 1 . $$

For the linear propagator (\(\delta = 1\)) the constraint on the exponent is slack. Absence of manipulation requires the level kernel to be positive definite as a quadratic form, and Alfonsi, Schied and Slynko (2012) show that a non-negative, decreasing, convex kernel is enough to exclude the transaction-triggered kind as well. Every power law with \(\gamma > 0\) qualifies.

Put the two together and they constrain the impact function, not the choice between readings: with \(\gamma = (1-\alpha)/2\), Gatheral’s inequality requires \(\delta \ge (1 + \alpha)/2\), which is 0.75 at \(\alpha = 0.5\). That sits above the square-root exponent of about 0.5 measured on metaorders, a tension that is one of the reasons nonlinear and latent-liquidity models exist.

What survives. No-arbitrage arguments carve out an admissible region in \((\gamma, \delta)\) and in kernel shape. They do not collapse the equivalence class. The transient model with the fine-balance kernel is admissible. The permanent model, with linear permanent impact on the unpredictable part of the flow, is admissible too: linear permanent impact is the textbook manipulation-free case (Huberman and Stanzl, 2004). Both reproduce the tape. Selecting between them needs information that is not in unconditional sign and return covariances:

  • Labelled metaorders. With parent-order identifiers, start and end times, the impact path during execution and the reversion after it are measured directly, conditional on the object you actually trade.
  • Event-level book data. Taranto et al. show that TIM and HDIM coincide with a single event type and separate once events are split, for example into price-changing and non-price-changing trades, or once queue depletion and refill are observed. L3 data makes the liquidity response itself an observable.
  • Exogenous variation. Randomized child-order timing or sizing in your own flow is the one source of signs that is uncorrelated with information by design.

Three checks before deploying an estimated kernel

None of these makes public tape sufficient. They tell you how much of your kernel is evidence about impact and how much is a restatement of the sign autocorrelation.

1. Participation-regime stability, against a signs-only benchmark

Re-estimate \(G\) on subsamples where order splitting differs: index-rebalance closes and expiry days against ordinary sessions, the opening half hour against midday, high against low institutional participation. A structural kernel should not move when \(C(\ell)\) moves. A forecast-error filter must.

Then run the sharper version. In each subsample, compute the forecast weights \(a_k\) from \(C(\ell)\) alone, with no prices, and compare \(-a_k\) with the estimated \(G(k)/G(0)\) lag by lag. If the price-free benchmark reproduces the kernel shape within estimation error, the price data contributed one number, \(G(0)\), and the decay profile contains no independent information about impact. Report the fraction of the kernel’s variation across lags that the benchmark explains. Anything near one means the scheduler is being parameterized by order-flow persistence.

2. Cross-impact and index-arbitrage leakage

A single-asset fit attributes every sign-correlated price move to the asset’s own trades. Index futures and ETF arbitrage create flow in the stock that is correlated with moves in the basket, and basket moves are persistent at exactly the horizons where the tail of the kernel is estimated. The result is a slow component in \(\bar G\) that is common-factor price discovery, not self-impact.

Before deconvolution, project \(\Delta P_t\) on index-futures and sector-ETF returns and fit the kernel on the residual, then compare with the raw fit. Two cautions. Contemporaneous residualization also removes genuine own impact for names that are large in the index, so repeat with strictly lagged factor returns as a bound. And the cleaner design, where the data exist, is a multivariate propagator with the futures’ signed flow as an extra regressor, which is the cross-impact setting. If the long-lag kernel changes materially between the raw and residualized fits, the tail is not yours to schedule against.

3. Reconciliation against metaorder implementation shortfall

The only test with authority is out of sample and on your own parent orders. For each historical metaorder, push the realized child schedule through the estimated kernel to get a predicted impact path during execution and a predicted reversion path afterwards. Compare with realized slippage against arrival, bucketed by participation rate and duration, and with post-completion marks at fixed multiples of the execution duration.

Three signatures are worth reading:

  • Predicted reversion exceeds realized reversion. The kernel’s \(\bar G(\infty) \approx 0\) is an artifact of the forecast weights summing to one. Treat part of the impact as permanent and re-fit that share on your own data.
  • Shortfall depends on total size and barely on the schedule. The schedule-dependent savings your optimizer reports are a product of assumed relaxation. The simpler square-root estimate may forecast cost better.
  • Errors sort by how detectable the order was. If high-participation, regular schedules come in cheaper per unit than the kernel predicts, and irregular ones dearer, that is the history-dependent mechanism showing through.

Fit to \(\mathcal{R}(\ell)\) is not a validation step. It is satisfied by construction for every member of the equivalence class.

Production failure modes

  • Reading uniqueness as identification. The Toeplitz system always has one solution. That is a property of linear algebra, not of the market.
  • Calibrating resilience from the tape. An Obizhaeva–Wang resilience rate backed out of a public-tape kernel inherits the forecast-filter shape, and the block-trade structure of the optimal schedule inherits the error.
  • Truncation that hides the tail. With long memory, cutting the system at \(L\) lags biases the long-lag coefficients, and the bias depends on \(L\). Check that conclusions hold as \(L\) doubles.
  • Trade-sign misclassification. Signing errors attenuate \(C(\ell)\) and \(\mathcal{R}(\ell)\) differently and change the implied forecast weights. Odd-lot and off-exchange prints deserve their own check.
  • Trade time against clock time. The kernel is estimated per trade. A scheduler in clock time needs a volume-clock mapping that is itself regime dependent.
  • One kernel across the capitalization range. Sign persistence differs by tick size, queue length and venue mix, so the kernel does too, for the reasons above.

Questions to ask a paper (or a vendor’s cost model)

  1. Is the kernel estimated from public tape, from labelled metaorders, or from event-level book data?
  2. Does the paper report how much of the kernel is reproduced by the sign autocorrelation alone?
  3. Is impact decay tested against a null model with permanent impact, or only fitted?
  4. Is the kernel shown to be stable when order-flow persistence changes?
  5. Are common-factor moves removed, and how, before the single-asset fit?
  6. Is the validation an out-of-sample comparison with realized implementation shortfall, or a fit to the response function?
  7. Do the fitted exponents satisfy the relevant no-manipulation condition for the impact function actually used?

Related reading in the corpus: the no-go theorem itself, a queue-sensitive propagator for passive impact, offline learning of propagator models, identifying metaorders from public data, and generating realistic metaorders from public data. The topic hubs are market microstructure and HFT & execution.


Walk the argument: the paper’s flowchart and evidence card → · Run your numbers: transaction-cost break-even calculator → · capacity estimator → · The foundations: transaction costs, slippage and market impact → · how to evaluate a microstructure paper → · Hawkes processes and order flow →

Sources and background: Hasbrouck (1991) for the trade-innovation view of permanent impact; Bouchaud, Gefen, Potters and Wyart (2004) for the propagator model and the fine-balance exponent; Lillo and Farmer (2004) for long-memory order flow with adjusting liquidity; Huberman and Stanzl (2004) and Gatheral (2010) for price manipulation and no-dynamic-arbitrage; Alfonsi, Schied and Slynko (2012) for kernel conditions; Obizhaeva and Wang (2013) and Almgren and Chriss (2001) for the execution models; Taranto, Bormetti, Bouchaud, Lillo and Tóth (2018) for transient against history-dependent impact; Fujiwara, Sato and Kanazawa (2026, arXiv 2611.10574) for the no-go theorem. The discrete derivation and the worked table are our own. Browse the flowchart library, the guides, and the blog.