Persistence, patience and costly information acquisition

Benjamin Davies

arxiv: 2603.11453 · v2 · submitted 2026-03-12 · 💰 econ.TH

Persistence, patience and costly information acquisition

Benjamin Davies This is my paper

Pith reviewed 2026-05-15 12:35 UTC · model grok-4.3

classification 💰 econ.TH

keywords information acquisitionAR(1) processbelief precisionpersistencepatiencecostly signalsdynamic optimization

0 comments

The pith

A forward-looking agent optimally balances signal costs against belief accuracy for a persistent Gaussian state, with patience yielding sharper beliefs at lower total expense.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper studies how an agent who cares about future payoffs gathers costly signals about a state that follows an AR(1) process. It derives the optimal policy for choosing how much precision to buy each period and then traces the resulting long-run belief variance and cumulative costs. Greater patience tightens beliefs and reduces total spending; greater persistence raises costs and affects precision in a non-monotonic way. Readers care because the setting covers many repeated decision problems in which information is valuable yet expensive.

Core claim

The agent solves a dynamic program that trades off the immediate cost of signal precision against the value of lower future belief variance. In steady state the chosen precision is a function of the AR(1) coefficient and the discount factor; higher persistence produces a hump-shaped pattern in belief precision while always increasing average costs, whereas higher patience produces both higher precision and lower average costs.

What carries the argument

The optimal information-acquisition policy, which maps current posterior variance, the persistence parameter, and the discount factor into the precision of the signal the agent chooses to purchase.

If this is right

Belief precision rises and total information costs fall as the agent's discount factor increases.
Higher persistence always raises steady-state information costs even though its effect on belief precision is non-monotonic.
The optimal policy can be computed recursively from the current belief variance and the two parameters.
In the long run the agent settles into a stationary distribution of belief variance whose mean depends only on persistence and patience.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same trade-off could be used to study how patient central banks decide how much to spend on economic data collection.
If multiple agents observe the same state, their joint policy might generate strategic complementarities in information acquisition.
Numerical solutions for non-Gaussian or nonlinear state processes would show whether the non-monotonicity in precision survives.

Load-bearing premise

The unobserved state follows a Gaussian AR(1) process and the cost of acquiring a signal depends only on its precision and is separable from the belief-update rule.

What would settle it

A laboratory experiment in which subjects face repeated draws from an AR(1) process and can buy signals of chosen precision; the data would falsify the claim if average costs do not rise with the estimated persistence parameter or if precision does not increase with the imposed discount factor.

read the original abstract

A forward-looking agent observes signals of a state that follows a Gaussian AR(1) process. He balances the cost of having imprecise beliefs with the cost of acquiring more precise signals. I characterize his optimal information acquisition policy, and analyze how his steady-state beliefs and costs depend on persistence (the AR(1) parameter) and patience (the agent's discount factor). Higher persistence has a non-monotone effect on belief precision and raises overall costs. Higher patience makes beliefs more precise and lowers overall costs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Explicit characterization of optimal policy with non-monotone persistence effects in costly dynamic information acquisition.

read the letter

The main thing to know is that this paper delivers an explicit characterization of the agent's optimal signal-precision choice in a setting where the unobserved state follows a Gaussian AR(1) and the agent pays a cost for better signals. It then derives how the long-run belief variance and the total information costs vary with the AR(1) coefficient and with the discount factor. Higher persistence raises overall costs but affects belief precision in a non-monotone way; higher patience produces more precise beliefs and lower total costs. These comparative statics come from solving the dynamic program on the variance state variable and then differentiating the resulting steady-state fixed point. The Gaussian structure keeps everything one-dimensional and tractable, which is what lets the author get past the usual qualitative statements in this literature. The stress-test note confirms the recursion is deterministic given the chosen precision and that the derivatives follow without circularity or unstated boundary problems. The construction therefore looks internally consistent on its own terms. The main limitation is scope rather than a flaw in the argument: everything rests on separability of the signal cost from the belief process and on the state remaining Gaussian. Relax either and the non-monotonicity could disappear, but the paper does not claim otherwise. The free parameters are just the persistence coefficient, the discount factor, and the cost function, which is standard. This is the sort of paper that researchers building macro or finance models with endogenous information choice will find immediately usable for calibration exercises or for varying persistence across series. It supplies a concrete comparative-statics device instead of another existence result. Because the central claims are derived directly from the primitives and the derivation checks out, the paper is worth sending to referees for a full evaluation rather than a desk rejection.

Referee Report

0 major / 2 minor

Summary. The paper models a forward-looking agent who acquires costly signals about a state that follows a Gaussian AR(1) process. The agent trades off the cost of imprecise beliefs against the cost of more precise signals. The author characterizes the optimal dynamic information-acquisition policy via dynamic programming and derives the dependence of steady-state belief variance and total costs on the AR(1) persistence parameter and the agent's discount factor. Higher persistence produces a non-monotone effect on belief precision and raises overall costs; higher patience produces more precise beliefs and lowers overall costs.

Significance. If the derivations hold, the paper supplies a clean, tractable framework for dynamic rational inattention with persistent shocks. The Gaussian structure yields an explicit variance recursion whose fixed point can be differentiated implicitly, delivering unambiguous comparative statics on persistence and patience. These results are directly relevant to models of learning, forecasting, and costly attention in macroeconomics and finance.

minor comments (2)

The abstract does not mention the dynamic-programming approach or the fixed-point construction of the steady-state variance; adding one sentence would improve accessibility.
Notation for the signal-precision choice variable and the resulting posterior variance should be distinguished more clearly in the main text (e.g., by using distinct symbols).

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of the paper and the recommendation for minor revision. The referee's summary correctly identifies the core contribution: the characterization of the optimal dynamic information-acquisition policy for a persistent Gaussian AR(1) state and the comparative statics on persistence and patience. We are pleased that the tractability of the Gaussian setup and its relevance to macro and finance models are noted.

Circularity Check

0 steps flagged

No significant circularity in the derivation

full rationale

The paper's central result is a recursive characterization of the optimal signal-precision policy obtained via dynamic programming on the agent's value function. The steady-state belief variance is recovered as the fixed point of the explicit Gaussian AR(1) prediction-error recursion composed with the optimal precision choice; comparative statics with respect to the AR(1) persistence parameter and the discount factor then follow from implicit differentiation of that fixed-point equation. No data-fitting step, no load-bearing self-citation, and no redefinition of inputs as outputs appear in the construction, which remains self-contained under the maintained assumptions of Gaussian dynamics and separable costs.

Axiom & Free-Parameter Ledger

3 free parameters · 2 axioms · 0 invented entities

The model rests on standard rational-expectations updating under quadratic costs and Gaussian signals; no new entities are introduced. The AR(1) persistence parameter and the per-signal cost function are primitives taken as given rather than fitted.

free parameters (3)

AR(1) persistence coefficient
Taken as an exogenous parameter whose effect on steady-state precision and costs is characterized.
Discount factor
Exogenous parameter whose effect on steady-state precision and costs is characterized.
Signal cost function
Functional form of the cost of signal precision is a modeling choice required for the optimization problem.

axioms (2)

domain assumption The agent is forward-looking and solves a dynamic optimization problem with discounting.
Standard in dynamic information-acquisition models; invoked to define the value function.
standard math Belief updating follows Bayes' rule under Gaussian signals and quadratic loss.
Required for the Kalman-filter-like updating that underlies the precision dynamics.

pith-pipeline@v0.9.0 · 5356 in / 1476 out tokens · 26800 ms · 2026-05-15T12:35:14.690374+00:00 · methodology

Persistence, patience and costly information acquisition

Core claim

What carries the argument

If this is right

Where Pith is reading between the lines

Load-bearing premise

What would settle it

discussion (0)