we-solve (beta)

Econ-592 · economics · Economics

Strategic adaptation after regime changes

LLM status Open unverified

Statement

Original subject group: Game theory and strategic interaction

Audit classification: Published research agenda. Fudenberg–Levine p.153 and conclusion discusses reevaluation and regime changes. The latent two-model experiment is editorial.

Definitions and precise research target

Model and research target. Let human players repeatedly interact in a finite game with known action sets. At an experimentally randomized date $\tau$, an institution changes the payoff matrix from $u^0$ to $u^1$. Players may receive an announcement or learn the change only through realized outcomes; this disclosure $Z$ is randomized by interacting group. A candidate learning model maintains latent state $M_{it}\in\{\text{old model},\text{revised model}\}$ and parameter estimates conditional on that state. Observe actions, realized payoffs, and incentivized predictive beliefs $b_{it}$. Define revision time $R_i$ by a prespecified belief-consistency criterion, rather than by an unobservable analyst judgment. The principal estimands are \[ \Pr(R_i-\tau\leq h\mid do(Z)) \quad\text{and}\quad E\left[\sum_{t=\tau}^{\tau+h}(u_i^{\mathrm{best}}(t)-u_i(a_t))\mid do(Z)\right], \] where $h$ is a fixed horizon and $u_i^{\mathrm{best}}(t)=\max_{a_i}u_i^1(a_i,a_{-i,t})$. Determine whether observations distinguish changing beliefs within one model from switching the model itself, and derive identified sets when they cannot. Estimate disclosure effects and validate revision triggers on new changes. Regret here is a finite experimental outcome, not an equilibrium-existence target. Group randomization and explicit signal measurement assumptions are necessary because opponents also adapt after the change.

Definitions and maintained assumptions (editorial unless a source definition is named). Fix finite actions, both payoff tensors, randomized change date $\tau$ and disclosure treatment $Z$, and whether $\tau$ is known. At each date record the whole action profile and each player’s prediction distribution over declared outcomes. The old-model and revised-model hypotheses must specify observable prediction laws; labels alone do not define statistically distinguishable latent states. For tolerance $\eta$ and window $w$, define $R_i=\inf\{t\ge\tau:\|b_{is}-b^{\rm revised}_{is}\|\le\eta\text{ for }s=t,\ldots,t+w-1\}$, with $+\infty$ if never satisfied; the benchmark forecast and norm are part of the protocol. The sum of realized best-response shortfalls compares the chosen action with the best action against opponents’ realized actions, not a feasible counterfactual policy that changes their behavior. Censoring at experimental termination is reported explicitly. Randomization is by interacting group, and measurements before disclosure cannot be conditioned on treatment-induced post-change variables without changing the estimand.

Variants and assumption changes

  1. Known change (editorial). Announce both payoff tensors and the date. Estimate response speed to information and distinguish it from slow payoff learning.
  2. Hidden change (editorial). Hide the changed tensor and provide noisy payoff feedback. Characterize observational equivalence between continuous parameter updating and discrete model replacement using additional randomized diagnostic outcomes.

References and source audit

  1. p.153 and conclusion pp.165–166. Supports the stated research area; exact displayed specification is editorial. https://economics.mit.edu/sites/default/files/2022-09/Whither%20Game%20Theory%20Towards%20a%20Theory%20of%20Learning%20in%20Games.pdf
  1. Drew Fudenberg; David K. Levine. 2016. Whither Game Theory? Towards a Theory of Learning in Games. Introduction, p.153; conclusion, pp.165–166. https://economics.mit.edu/sites/default/files/2022-09/Whither%20Game%20Theory%20Towards%20a%20Theory%20of%20Learning%20in%20Games.pdf. DOI: https://doi.org/10.1257/jep.30.4.151.

Common conventions: Source mathematical conventions

These conventions apply to each entry unless explicitly replaced.

Models and identification. A primitive model $m$ specifies all probability laws, feasible choices, information, equilibrium and measurement rules used by its target. The admissible class $\mathcal M$ is fixed independently of outcomes. If $P_O(m)$ is its observable probability law and $\tau(m)$ the target, its identified set at observable law $P$ is \[ \mathcal I_\tau(P)=\{\tau(m):m\in\mathcal M,\ P_O(m)=P\}. \] Point identification means that this set is a singleton. An empty set signals incompatibility of data and assumptions. Coordinate bounds need not describe the joint set. Bounds are sharp only if every retained target is attainable or arbitrarily approachable by compatible models. Finite bounds require explicit restrictions on unobserved quantities; unrestricted confounding, hidden wealth, extrapolation or arbitrary equilibrium selection can yield uninformative sets.

Causal interventions. A potential outcome $Y_i(p)$ is the outcome under a complete policy regime $p$, including timing, financing, information and market spillovers. The target population and units are fixed before treatment. With interference, use the complete assignment vector or a justified exposure mapping; individual no-interference notation alone cannot identify an economy-wide intervention. A difference of mean potential outcomes does not require the cross-world joint distribution. Conditioning on post-treatment migration, survival, enrollment or employment changes the estimand and needs separate justification.

Statistical statements. An estimator, confidence set, forecast or test specifies a sampling law, model class and loss/coverage criterion. Uniform $(1-\alpha)$-coverage means $\inf_{m\in\mathcal M}\Pr_m\{\tau(m)\in C_n\}\geq1-\alpha$, or an explicitly asymptotic version. The whole parameter space trivially has coverage, so informativeness requires a stated width, risk or power objective. A forecasting improvement is relative to a named benchmark, information set and evaluation population.

Equilibrium and optimization. An equilibrium comprises feasible plans, optimality, beliefs where needed, budget constraints and all market/resource clearing. The primitives or a stated benchmark define each correspondence $\mathcal E(p;m)$. Nonemptiness is assumed or proved, never inferred just from compact policy sets. A selected-equilibrium target uses a declared selection rule; a robust target quantifies over all permitted equilibria. Use suprema or infima unless attainment is established. An $\varepsilon$-optimal policy is feasible and within $\varepsilon$ of the finite optimum value. Report an empty feasible set separately.

Welfare and accounting. Normative utility, interpersonal normalization, social weights, numeraire, discounting and the use of public receipts are primitives. Behavioral utility need not coincide with the welfare criterion. Household welfare includes ownership income and rents once; adding profits again to compensating variation generally double-counts them. A monetary equivalent is defined by an explicit transfer and price convention, with monotonicity and range conditions. Average effects, mechanism contributions, transfers and welfare losses are different targets. Infinite sums require convergence and dynamic feasible plans require terminal or no-Ponzi conditions.

Source versions. A working-paper date, revision date and journal publication year are distinguished. A DOI for a journal version is not automatically the DOI of an earlier manuscript. Locators refer to the stated version and printed pages where specified. A metadata-only check does not verify a claim about a full-text conclusion. Every entry's source subsection narrows the provenance claim accordingly.

Statement licence: CC-BY-4.0 · Source file

Contributions 1

Econ-592-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-592

Strategic adaptation after regime changes: Diagnostic revision probabilities and censored belief windows

Two different meanings of revision

In a finite repeated game, a randomized payoff change occurs at date $\tau$, with disclosure randomized by interacting group. Revision is measured using a prespecified belief criterion. Writing $c_t=\mathbf1\{\|b_t-b_t^{\mathrm{revised}}\|\le\eta\}$, the registered window criterion is \[ R=\inf\{t\ge\tau:c_t=\cdots=c_{t+w-1}=1\}. \tag{1} \] An unobserved model-state switch is a different object. The call for learning theories with model reevaluation in [fl, pp.153,165--166] motivates both, but does not identify them with one another. We study a diagnostic design for a specified latent switch and derive exact bounds for the observable window target when follow-up ends too early.

Calibrated diagnostics for an absorbing state

Let $M_t\in\{0,1\}$ be a latent old/revised state that can move only from zero to one. A binary diagnostic $D_t$, recorded before that round's actions, has known state-dependent probabilities \[ \Pr(D_t=1\mid M_t=0,Z=z)=a_t, \qquad \Pr(D_t=1\mid M_t=1,Z=z)=b_t, \qquad a_t\ne b_t. \tag{2} \] These equations must hold after integrating the within-state histories and types in each disclosure arm. They are not supplied by naming a report "diagnostic." They may be justified by calibration, or instead treated as restrictions to test. Diagnostic administration is part of the intervention regime; if probes change subsequent learning, the target is learning under that probe protocol. Let $F_t(z)=\Pr(M_t=1\mid \operatorname{do}(z))$.
Theorem: With independent group randomization, consistency, positive arm probabilities, and (2), population diagnostic means identify \[ F_t(z)=\frac{E[D_t\mid Z=z]-a_t}{b_t-a_t}. \tag{3} \] If $F_{t-1}(z)<1$, the conditional revision hazard is identified as \[ \Pr(M_t=1\mid M_{t-1}=0,\operatorname{do}(z)) =\frac{F_t(z)-F_{t-1}(z)}{1-F_{t-1}(z)}. \tag{4} \] An error at most $\delta$ in the diagnostic mean creates an error at most $\delta/|b_t-a_t|$ in (3), when calibration values are exact.
Proof: Total probability and randomization give the mean $a_t(1-F_t)+b_tF_t$, whose nonzero slope gives (3) and the error bound. Absorption implies that the event $\{M_t=1\}$ is the disjoint union of $\{M_{t-1}=1\}$ and $\{M_{t-1}=0,M_t=1\}$. Subtract their probabilities and divide by $1-F_{t-1}$ to obtain (4). ∎
The restrictions imply $F_t\in[0,1]$ and nondecreasing $F_t$. Violations at the population level reject calibration or absorption. When $a_t=b_t$, the diagnostic is uninformative about this state; more data cannot repair the missing contrast. Uncertain $a_t,b_t$ instead produce a set of compatible sequences, with the same values required across every arm to which the calibration restriction applies.

Why a successful hidden-state fit does not establish a switch

Proposition: Any specified finite-state hidden Markov learning model has an observationally equivalent predictive model whose internal state is the continuously valued posterior distribution over those hidden states. Consequently observable actions, payoffs, and reports alone do not distinguish those two interpretations of internal representation.
Proof: For simplicity write the hidden states as $0,1$ and the next recorded observation as $Y_t$ in a finite alphabet. Include recorded actions and reports in that alphabet. If the current posterior is $\pi$, the specified transition matrix gives a predicted probability $\pi^-$ of state one. With emission probabilities $f_0(y),f_1(y)$, generate the next observation with probability $(1-\pi^-)f_0(y)+\pi^-f_1(y)$ and update the internal number to \[ \pi'= \frac{\pi^-f_1(y)}{(1-\pi^-)f_0(y)+\pi^-f_1(y)}. \tag{5} \] For zero-probability observations define the update arbitrarily. Use the known date and recorded history in this recursion whenever transitions or emissions are history dependent. Conditional on every positive-probability history, this continuous-state model has exactly the same next-observation law as the hidden-state model. Starting with the same initial law, induction on history length gives equality of all finite-dimensional observable laws. No latent discrete state needs to be sampled in the predictive construction. ∎
Thus (3) identifies the mixture weight of a declared calibrated state model; it does not prove that a person's mind implements discrete replacement rather than probabilistic updating within one larger model. Diagnostic interventions help discriminate restricted prediction laws, not unrestricted relabelings of an internal representation.

Sharp correction for the forward-looking window

Set $\tau=0$ by changing the time origin, and suppose $c_0,\ldots,c_T$ are observed completely. Fix $0\le H\le T$. Define $\ell_H=1$ if there is a start $s\le H$ with $s+w-1\le T$ whose whole observed window contains only ones. Define $u_H=1$ if there is a start $s\le H$ for which every observed entry of the window $s,\ldots,s+w-1$ is one. An incomplete window can qualify for $u_H$.
Theorem: Without restrictions on beliefs after $T$, the sharp bounds in each randomized disclosure arm are \[ E[\ell_H\mid Z=z]\le \Pr(R\le H\mid\operatorname{do}(z)) \le E[u_H\mid Z=z]. \tag{6} \] When $H+w-1\le T$, the two endpoints coincide.
Proof: A completely observed successful window already proves $R\le H$, and an observed zero in every candidate window makes that event impossible. This proves the pointwise lower and upper bounds. Complete all unobserved entries with zeros. A newly completed window cannot succeed, so the event is exactly $\ell_H=1$. Complete them instead with ones; precisely the windows without observed zeros succeed, so the event is exactly $u_H=1$. These completions preserve every observed history and attain the endpoint probabilities. Randomization identifies the observed arm-specific means. If every candidate window is completely observed, the definitions of $\ell_H,u_H$ are identical. ∎
Although the event is indexed by its window's starting date, it is only confirmed $w-1$ dates later. Treating $R$ as an ordinary stopping time at its starting date would conceal this censoring problem. The same all-zero and all-one completions attain the endpoint CDFs simultaneously for every $H\le T$.

Residual agenda

Whole-group randomization also identifies mean realized best-response shortfalls over observed horizons by ordinary arm comparisons; that shortfall is not a policy counterfactual with fixed opponent behavior. The remaining substantive questions are valid belief measurement, calibration of distinct model predictions, external validation on new payoff changes, and restrictions connecting latent switches to the observable criterion (1). None follows from fitting a two-state model.

Completion Estimate

56%

This is a post hoc, heuristic estimate for the full catalogue target. The attempt identifies calibrated absorbing-state revision probabilities, proves equivalence with continuous posterior updating, and derives sharp censoring bounds for observable belief windows. Valid belief measurement, credible diagnostic calibration, distinction between internal model interpretations, and external validation of disclosure and revision triggers on new strategic changes remain missing.

References

  • [fl] D. Fudenberg and D. K. Levine (2016), Whither Game Theory? Towards a Theory of Learning in Games, Journal of Economic Perspectives 30(4),151--170. https://doi.org/10.1257/jep.30.4.151.

Go to contribution →