---
title: 'No-Anticipation: Modeling Future-Info Restrictions'
url: https://www.emergentmind.com/topics/no-anticipation-assumption
type: topic
---

# No-Anticipation: Modeling Future-Info Restrictions

The **no-anticipation assumption** is a family of domain-specific restrictions that exclude current states, decisions, prices, or outcomes from depending on future information in a prohibited way. Its meaning is not uniform across fields. In some literatures it is a standard adaptedness or causality requirement; in others it is a benchmark assumption whose precise content depends on how interventions, filtrations, rollout schedules, or future event times are specified. Recent work shows that the phrase is often ambiguous unless the relevant information structure is stated explicitly, and that relaxing it can either expose hidden benchmark simplifications or define entirely new tasks and models [2507.12891][0804.2035][2606.04258].

## 1. Conceptual scope and formal ambiguity

A central theme in recent literature is that “no anticipation” is not a single theorem-like condition but a modeling choice whose content varies with the object being intervened on. In difference-in-differences, the standard verbal formulation—future treatment has no effect before implementation—can appear redundant in temporally ordered causal models. The ambiguity, according to "Refining the Notion of No Anticipation in Difference-in-Differences Studies" [2507.12891], comes from conflating an intervention on **policy implementation** with an intervention on the earlier **decision/plan/announcement** to implement the policy. In asynchronous systems, "The non-anticipation of the asynchronous systems" [0804.2035] reaches a parallel conclusion from a different direction: it introduces several inequivalent causal notions rather than a single universal definition. In decision-theoretic portfolio theory, "Anticipatory Portfolio Optimization" [2606.04258] treats no anticipation as the restricted benchmark in which the optimizer uses a base information set \(I_0\) and a fixed reference law, whereas the anticipatory controller solves under richer information \(I\) and the deployed law \(P_\theta\).

| Domain | Baseline no-anticipation restriction | Representative refinement |
|---|---|---|
| DiD | Future implementation should not affect pre-treatment outcomes | Distinguish implementation \(A\) from decision/plan \(P\) [2507.12891] |
| Asynchronous systems | Output must not depend on future input | Several nonequivalent definitions, including first-switch and history-based causality [0804.2035] |
| Portfolio choice | Restricted estimator uses \(I_0\) and fixed \(\bar P\) | Anticipatory optimizer uses richer \(I\) and \(P_\theta\) [2606.04258] |
| Finance | No sure profits from future-timed information | Equivalent to absence of predictable jumps [1708.03099] |

These formulations are technically different, but they share a common role: they specify which future-dependent structures are excluded from admissible reasoning. This suggests that no anticipation is best understood as an **information-structure constraint** rather than as a single substantive claim.

## 2. Sequential prediction, perception, and machine anticipation

In action anticipation, recent work shows that standard benchmarks often embed hidden future information. "Untrimmed Action Anticipation" [2202.04132] identifies a benchmark assumption that test-time clips are sampled at a fixed offset before the start of a labeled action. In the trimmed formulation, the model predicts the class \(y\) of an action starting at time \(\tau_s\) from the segment beginning at \(\tau_s-(\tau_o+\tau_a)\) and ending at \(\tau_s-\tau_a\). This effectively tells the model *when* the next action will begin up to the fixed anticipation time \(\tau_a\). The proposed untrimmed task removes that assumption: at timestamps \(t_1,\dots,t_n\), sampled every \(\alpha\) seconds, the model predicts the set \(y_a=\{(c_i,\tau_i)\}_{i=1}^{N_a}\) of actions beginning within a horizon \(T\). The EPIC-KITCHENS-100 conversion uses \(\alpha=0.25\) seconds and \(T=5\) seconds, yielding 1,057,238 training timestamps and 185,532 validation timestamps; 38% of timestamps have no future action in the horizon and 36% contain at least two future action labels. Adapted trimmed baselines remain weak: on action anticipation, RU obtains \(0.84\) mAP\(_{1.0}\) and \(2.05\) mAP\(_{\inf}\), while RU-5-clf reaches \(0.87\) mAP\(_{1.0}\) and \(2.55\) mAP\(_{\inf}\), with false positives on 98% of no-action timestamps for the best model. The paper therefore reframes anticipation as a form of future temporal detection requiring answers to *whether*, *what*, and *when*.

The surgical anticipation literature removes a closely related assumption. "Rethinking Anticipation Tasks: Uncertainty-aware Anticipation of Sparse Surgical Instrument Usage for Context-aware Assistance" [2007.00548] criticizes setups in which a new action is assumed to occur within typically one second and only its category is unknown, or in which dense temporal segmentations are required during training and inference. Instead, it defines, for each frame \(x\) and instrument \(\tau\), a truncated remaining-time target
\[
r_h(x,\tau)=\min\{t_x(\tau),h\},
\]
with \(h\in\{2,3,5,7\}\) minutes, together with an auxiliary three-way label \(c_h(x,\tau)\in\{\text{anticipating},\text{present},\text{background}\}\). At inference the method uses only image data, without dense action segmentation, phase labels, instrument presence annotations, future temporal boundaries, or knowledge of final surgery duration. This is a strict no-future-timing formulation at test time, even though target construction uses future occurrences offline during training.

Other recent work questions whether no anticipation should be identified with dense temporal video processing. "Understanding Multimodal Complementarity for Single-Frame Action Anticipation" [2601.22039] studies prediction of the action occurring \(\delta=1s\) ahead from a single RGB frame, optionally depth, and semantic history. The resulting AAG+ framework improves the original AAG and, on IKEA-ASM, reaches \(51.26/88.88\) top-1/top-5 compared with VLMAH at \(52.31/85.44\); on Assembly101, however, AAG+ remains below AVT and VLMAH on Recall@5, indicating that dense temporal context still matters in higher-ambiguity settings. The paper therefore challenges the assumption that explicit temporal video input is always necessary, but not the need for structured context.

A different boundary appears in long-term multimodal anticipation. "Multi-level and Multi-modal Action Anticipation" [2506.02382] operates on partially observed video prefixes, yet its segmentation and multimodal transformers use unmasked self-attention and cross-attention, so it is best characterized as prefix-based at the sequence level rather than strictly causal within the prefix. By contrast, "Zero-Shot Anticipation for Instructional Activities" [1812.02501] keeps the test-time constraint explicit: given ingredients and an observed text or video prefix, it predicts subsequent steps before seeing those segments, using external instructional corpora to learn procedural structure offline.

## 3. Causal inference and experimental design

In causal inference, the modern literature increasingly treats no anticipation as a question about the *correct intervention*, not merely about temporal ordering. In a two-period DiD setup, "Refining the Notion of No Anticipation in Difference-in-Differences Studies" [2507.12891] notes that standard DiD identifies
\[
ATT_{A_2}=\mathbb{E}(Y_2^{a_2=1}-Y_2^{a_2=0}\mid A_2=1)
\]
under positivity, consistency, and parallel trends formulated with respect to \(A_2\), without any extra no-anticipation restriction. The paper then enlarges the model with a decision variable \(P\), where \(A_2^{p=p^*}=p^*\), and redefines no anticipation as
\[
\mathbb{E}(Y_1^{p=1}\mid P=1)=\mathbb{E}(Y_1^{p=0}\mid P=1).
\]
Under parallel trends with respect to \(P\), consistency for \(P\), and an exclusion restriction on direct effects of \(P\) on \(Y_2\), the classic DiD contrast equals \(ATT_{A_2}-\psi\), where
\[
\psi=\mathbb{E}(Y_1^{p=1}\mid P=1)-\mathbb{E}(Y_1^{p=0}\mid P=1).
\]
Under Assumptions 5–7, the same contrast identifies
\[
ATT_P=\mathbb{E}(Y_2^{p=1}-Y_2^{p=0}\mid P=1).
\]
The paper’s main conceptual claim is that the usual subject-matter concern is typically about \(P\), not \(A\).

"A Joint Analysis of Sensitivity to Anticipation and Parallel Trends Violations" [2603.00868] pushes this further by showing that observed pre-trends confound anticipation and parallel-trends failure. Using
\[
\delta_t=\mathbb{E}[Y_t(0)-Y_{t-1}(0)\mid X=1]-\mathbb{E}[Y_t(0)-Y_{t-1}(0)\mid X=0]
\]
for untreated-trend deviations and
\[
\varphi_t=\mathbb{E}[Y_t(1)-Y_t(0)\mid X=1]
\]
for anticipation effects, the paper derives
\[
\mathrm{ATT}_1=\theta_1+\varphi_0-\delta_1
\]
and
\[
\Delta_s=\delta_s+\varphi_s-\varphi_{s-1}.
\]
Thus a pre-treatment event-study coefficient \(\Delta_s\) is not purely evidence about parallel trends unless \(\varphi_s=0\). The paper parameterizes anticipation increments as \(A_s=\varphi_s-\varphi_{s-1}\) with \(A_s\in[\underline A_s,\overline A_s]\), or alternatively as \(A_s=p_s^\star\Delta_s\) with \(p_s^\star\in[\underline p,\overline p]\), and derives sharp identified sets for \(\mathrm{ATT}_1\). It also shows that the identified set under joint deviations can be shorter than the set under parallel-trends violations alone only when anticipation is constrained away from zero; if \(0\in[\underline A_0,\overline A_0]\), that shortening cannot occur.

In stepped wedge cluster randomized trials, no anticipation is embedded directly in the analysis model. "On Anticipation Effect in Stepped Wedge Cluster Randomized Trials" [2504.08158] shows that the standard Hussey–Hughes model
\[
Y_{ijk}=\mu+\beta_j+\delta Z_{ij}+\alpha_i+\epsilon_{ijk}
\]
and the exposure-time interaction model
\[
Y_{ijk}=\mu+\beta_j+\delta(s_{ij})Z_{ij}+\alpha_i+\epsilon_{ijk}
\]
both encode the restriction that treatment-related effects begin only when \(Z_{ij}=1\). The anticipation-augmented forms add \(\gamma A_{ij}\):
\[
Y_{ijk}=\mu+\beta_j+\gamma A_{ij}+\delta Z_{ij}+\alpha_i+\epsilon_{ijk}
\]
and
\[
Y_{ijk}=\mu+\beta_j+\gamma A_{ij}+\delta(s_{ij})Z_{ij}+\alpha_i+\epsilon_{ijk}.
\]
If the true model is HH-ANT but HH is fit, then
\[
E(\widehat\delta_\phi)=\delta+\omega_{HH,\gamma}^{HH-ANT}\gamma,\qquad
\omega_{HH,\gamma}^{HH-ANT}=-\frac{6(1+\phi Q)}{(Q+1)(2+\phi Q)},
\]
so omission of anticipation biases the treatment estimate. The paper also proves \(V_{HH}(\widehat\delta)<V_{HH-ANT}(\widehat\delta)\) and \(V_{ETI}(\widehat\Delta)<V_{ETI-ANT}(\widehat\Delta)\), implying that modeling anticipation increases variance and that ignoring it in sample-size planning can lead to underpowered trials.

## 4. Finance, arbitrage, and anticipatory value

In mathematical finance, no anticipation is often formalized through admissible information and the exclusion of exploitable predictable price moves. "On the existence of sure profits via flash strategies" [1708.03099] studies a càdlàg adapted gains process \(X\) with jump process \(\Delta X_t=X_t-X_{t-}\). Its central result is that there are no sure profits via flash strategies if and only if \(X\) does not exhibit predictable jumps, and no constant profits via flash strategies if and only if \(X\) does not exhibit fully predictable jumps. Economically, this means that price moves at predictable dates may occur, but their direction cannot be known in advance if one wants to exclude sure high-frequency arbitrage; if the size is also known, even stronger arbitrage results. The paper therefore concludes that price changes at scheduled dates should only be due to unanticipated information releases.

"Convergence of the financial value of weak information for a sequence of discrete-time markets" [2205.05133] relaxes the no-anticipation benchmark more conservatively. Trading strategies remain predictable:
\[
\Theta_m^i\in\mathcal H_{m-1}\quad (m\ge 1),
\]
and in continuous time the admissible class \(\mathcal A_{\mathcal F(S)}\) consists of \(\mathcal F\)-predictable processes. The relaxation occurs at the level of beliefs: instead of revealing the exact future signal \(Y\), the agent knows only its law \(\nu\). The associated minimal probability measure is
\[
P^\nu(A)=\int_{\mathcal P}P[A\mid Y=y]\nu(dy),
\]
with \(dP^\nu=\xi(Y)\,dP\), \(\xi=d\nu/dP_Y\). The financial value of weak information is then
\[
u(x,\nu)=\inf_{Q\in\mathcal E^\nu}\sup_{\Theta\in\mathcal A_{\mathcal F(S)}}E_Q\!\left[U\!\left(x+\int_0^T\Theta_u\cdot dS_u\right)\right].
\]
This is a relaxation of no anticipation in the agent’s information set, but not in the admissibility of controls.

The finance literature also distinguishes between the *possibility* of anticipation and its *equilibrium expression*. "Does the Market Anticipate? Can it? Should it?" [2603.02187] studies a continuous-time setting with pre-horizon risk resolution. The model allows public predictive data \(D_t\) and posterior beliefs \(\pi_t^+\), so pre-disclosure anticipation is formally possible. Yet the paper argues that optimized trading can suppress anticipation of predictable risk outcomes. In particular, with relative information intensity
\[
RII_t:=\frac{(\sigma_t^{lD})^2}{\check r_t},
\]
low-RII means \(RII_t\le 1\), and under that condition the heterogeneous-market clearing price can have a steady-state **Status Quo Bias**, with \(\Pi_t^{\mathbf 1}\ll \tilde\pi_t^{\mathbf 1}\). The implication is that underreaction to predictable risks need not contradict no-arbitrage.

"Anticipatory Portfolio Optimization" [2606.04258] generalizes the benchmark across insider information, dynamic planning, and market impact. The restricted policy is
\[
\theta_{\mathrm{na}}\in\arg\max_{\theta\in\Theta}E_{r\sim P_{\theta_{\mathrm{na}}}}[U(\theta;r)\mid I_0],
\]
whereas the anticipatory policy is
\[
\theta_{\mathrm{an}}\in\arg\max_{\theta\in\Theta}E_{r\sim P_\theta}[U(\theta;r)\mid I].
\]
For permanent linear impact, the price-taking allocation is
\[
\theta_{\rm na}=(\Lambda+\gamma\Sigma)^{-1}\mu,
\]
while the impact-aware allocation is
\[
\theta_{\rm an}=(2\Lambda+\gamma\Sigma)^{-1}\mu.
\]
The value of anticipation is the realized control gap \(V=J(\theta_{\rm an})-J(\theta_{\rm na})\). The paper further shows that correctly specified anticipation creates value, vacuous anticipation has zero value, and misspecified anticipation is harmful, with exact penalty
\[
\frac12\operatorname{tr}(H^{-1}\Sigma_\varepsilon).
\]

## 5. Control, optimization, asynchronous systems, and collective dynamics

Outside finance, no anticipation often appears as a control-theoretic or systems-theoretic causality restriction. "Online Proactive Multi-Task Assignment with Resource Availability Anticipation" [2310.02353] is explicit that it adopts no anticipation only on the demand side. The method assumes no information on upcoming requests, no arrival distribution, and no assumption concerning future locations of requests; however, it does anticipate which currently busy agents will become available within a receding horizon \(H(k)=k\delta\). The resulting setting is therefore partial anticipation: demand-side no anticipation, supply-side anticipation.

In stochastic analysis, "Infinite Anticipation Backward Stochastic Differential Equations" [2511.15548] removes the standard no-future-dependence structure from the BSDE generator while retaining adapted solutions. The new IABSDE takes the form
\[
\left\{
\begin{aligned}
Y_t&=\xi_T+\int_t^T f\left(s,\{Y_r\}_{r\in[s,\infty)},\{Z_r\}_{r\in[s,\infty)}\right)ds-\int_t^T Z_s\,dW_s,\quad t\in[0,T],\\
Y_t&=\xi_t,\quad t\in[T,\infty),\\
Z_t&=\eta_t,\quad t\in[T,\infty).
\end{aligned}
\right.
\]
Under a path-Lipschitz condition (H1) and integrability condition (H2), the paper proves existence and uniqueness of an \(\mathcal F_t\)-adapted solution in \(\mathcal S_{\mathcal F}^2\times\mathcal M_{\mathcal F}^{2,\beta}\), and derives a comparison result. The paper interprets this as a dual counterpart to SDEs with infinite delay, showing that standard no-anticipation in the driver is not mathematically necessary if adaptedness of the generator value and suitable weighted estimates are preserved.

In classical systems theory, "The non-anticipation of the asynchronous systems" [0804.2035] develops several inequivalent formalizations for asynchronous circuits. The simplest one says that if \(u\) and \(x\) are both variable, then
\[
\min\{t\mid u(t-0)\neq u(t)\}\le \min\{t\mid x(t-0)\neq x(t)\},
\]
so the first output switch cannot precede the first input switch. A stronger, history-based version requires
\[
u_{|(-\infty,t)}=v_{|(-\infty,t)}
\Rightarrow
\{x_{|(-\infty,t]}\mid x\in f(u)\}
=
\{y_{|(-\infty,t]}\mid y\in f(v)\}.
\]
The paper then adds bounded-memory and delayed-window variants, and proves structural results such as preservation under subsystem restriction and time-shift normalization.

A distinct use of anticipation appears in collective-motion models. "Anticipation induces polarized collective motion in attraction based models" [1710.05692] contrasts the baseline assumption—each particle updates from current neighbor positions or headings—with a model in which the local center of mass is computed from anticipated positions. In the synchronous Local Attraction Model without anticipation, the regime diagram is: no group for \(c<0.2\), mills for \(0.2<c<1.7\), and swarms for \(c>1.7\). Adding anticipation causes cohesive polarized groups to emerge for \(c<0.2\) and again for sufficiently large \(c\), despite the model having no explicit alignment term. In this setting, relaxing the no-anticipation assumption qualitatively changes the phase behavior rather than merely refining estimation.

## 6. Cross-domain implications and open directions

Across these literatures, relaxing no anticipation rarely amounts to a uniform weakening of causality. More often, it exposes a hidden benchmark assumption that had already encoded privileged alignment, timing, or information. In action anticipation, knowing that evaluation clips are taken exactly \(\tau_a\) seconds before onset turns prediction into a materially easier task than continuous-video forecasting [2202.04132]. In DiD, an apparently standard no-anticipation formula may be either redundant or misdirected unless the intervention on \(A\) is separated from the earlier decision variable \(P\) [2507.12891]. In finance, predictable dates are admissible, but predictable jump direction or magnitude is not if flash-strategy arbitrage is to be excluded [1708.03099].

The literature also shows that “more anticipation” is not uniformly better. Single-frame action anticipation can be competitive when procedural state and semantic history are highly constraining, but its advantage weakens in more ambiguous, high-variability settings [2601.22039]. Anticipatory portfolio optimization yields positive value under correct specification, zero value under vacuous enrichment, and a precise overfitting penalty under misspecification [2606.04258]. Market microstructure models even suggest that optimized trading may suppress anticipation rather than express it immediately [2603.02187]. This suggests that the substantive question is not whether a model is anticipatory in the abstract, but which future-dependent structures are *admissible, identifiable, and decision-relevant* in a given domain.

Open problems differ by field but are structurally similar. Untrimmed video prediction highlights the need to model no-action periods, multiple future events, and accurate time-to-action estimation [2202.04132]. DiD sensitivity analysis shows that pre-trends need joint treatment as mixtures of anticipation and untreated-trend deviations rather than as one-dimensional diagnostics [2603.00868]. Stepped wedge trials require models that accommodate anticipation together with exposure-time heterogeneity and updated power calculations [2504.08158]. Stochastic analysis now admits infinite anticipation in BSDE generators, but only under strong measurability, Lipschitz, and integrability conditions [2511.15548]. Taken together, these works indicate that the no-anticipation assumption remains indispensable as a benchmark, but only after its information structure, intervention target, and admissibility role have been specified with technical precision.

Source: https://www.emergentmind.com/topics/no-anticipation-assumption