Information Divergence and Endogenous Evolution of Asymmetric Information

arXiv:2610.11200 · econ.TH · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Information Divergence and Endogenous Evolution of Asymmetric Information".

Warren: Here's the gist of the paper. Standard theories trace information asymmetry to privately held information, heterogeneous signals, or costly information acquisition <ref:2610.11200#pg6> The paper identifies a distinct mechanism: otherwise identical rational agents observing a common Markov state at asynchronous times can hold different predictive distributions because their information differs in age <ref:2610.11200#pg10> When the transition kernel separates information ages, independent Bernoulli observation arrivals make such divergence recur infinitely often almost surely, even with identical arrival probabilities <ref:2610.11200#pg10> When agents instead choose costly updating intensities, independent update realizations can regenerate heterogeneous information ages among ex ante identical agents <ref:2610.11200#pg10> Mechanism of Information Divergence The core mechanism identified is information age, defined as the time elapsed since an agent's most recent observation <ref:2610.11200#pg9> Together, the last observed state and its age determine the agent’s predictive distribution over the current state <ref:2610.11200#pg9> When transition kernel is age-separating—so observations of different ages generate distinct predictive distributions—asynchronous observation produces what we call information divergence <ref:2610.11200#pg10> Under independent, nondegenerate Bernoulli arrivals, this divergence recurs infinitely often almost surely, even when all agents face the same arrival probability <ref:2610.11200#pg10> Endogenous Updating and Staleness The framework endogenizes information updating by allowing agents to choose the intensity of a stochastic observation process <ref:2610.11200#pg10> When updating is costly, agents trade off the resource cost of maintaining a higher refresh rate against the expected loss from acting on stale information <ref:2610.11200#pg10> In a stationary Gaussian AR(1) environment, the consequences depend not only on common staleness as well as the age gap <ref:2610.11200#pg9> Holding the age gap fixed, greater common staleness raises each agent's predictive uncertainty but reduces the disagreement between their predictive forecasts <ref:2610.11200#pg9> This generates stale consensus: low disagreement despite high uncertainty <ref:2610.11200#pg9> Consequences of Common Staleness Common staleness also shapes incentives to disclose information age, as the value of certifying relative freshness declines as both observations become stale <ref:2610.11200#pg10> With multiple ordered information ages, costly certification yields a threshold equilibrium in which sufficiently fresh agents certify their exact age while sufficiently stale agents pool <ref:2610.11200#pg10> Under geometric information ages, more frequent updating weakly lowers the smallest and largest equilibrium disclosure cutoffs wherever positive-certification equilibria exist <ref:2610.11200#pg10> This stale consensus effect explains why agreement need not indicate better information: agents may issue similar forecasts because both have current information or because both rely on sufficiently stale observations <ref:2610.11200#pg9> Linear-Gaussian Specialization and Economic Implications In the linear Gaussian AR(1) specialization, predictive uncertainty depends on three factors: environmental volatility (sigma 2), state persistence (rho), and information age (Ai(t)) <ref:2610.11200#pg10> For a given sigma squared and rho!= 0, predictive uncertainty increases with Ai(t) as unobserved innovations accumulate <ref:2610.11200#pg10> Under the random walk process, predictive uncertainty about the current payoffrelevant state increases linearly with Ai(t), without bound <ref:2610.11200#pg10> This stale-consensus effect implies that disagreement can understate uncertainty precisely when the underlying information is old <ref:2610.11200#pg9> Applications and Conclusions The model provides a structural reason why forecast disagreement and uncertainty can move in opposite directions <ref:2610.11200#pg9> This stale-consensus result implies that agreement across institutions may reflect common staleness rather than precise current information <ref:2610.11200#pg10> The framework also bears on information-sharing policy: open banking and standardized observation timestamps could reduce age-based asymmetries and certification costs <ref:2610.11200#pg10> The staleconsensus result likewise cautions against treating agreement across institutions as evidence of accurate assessments <ref:2610.11200#pg9> The model also yields a testable implication: holding observed forecast dispersion fixed, forecast errors should be larger when the information underlying those forecasts is older on average <ref:2610.11200#pg9> Page 3 Information divergence occurs at date t if there exist agents i ≠ j such that their predictive distributions over the current state differ: mu i(t) ≡ Pr(X t ∈ ⋅∣ ∣ I i(t)) ≠ Pr (X t ∈ ⋅∣ ∣ I j(t)) ≡ mu j(t) <ref:2610.11200#pg9> Different information ages need not, in general, imply different predictive distributions; Definition 4 states the condition under which they do <ref:2610.11200#pg9> Age-Separating Transition Kernel The transition kernel Q is age-separating if for any a ≠ a' and x, x' ∈ X, Q a(∙∣ x) ≠ Q a'(∙∣ x') <ref:2610.11200#pg9> Thus observations of different ages induce distinct predictive distributions, regardless of the realized values of the most recent observations <ref:2610.11200#pg9> Dynamic Emergence and Persistence Theorem 1 isolates our central mechanism: asynchronous observation of an evolving payoffrelevant state can generate asymmetric information among otherwise identical rational agents <ref:2610.11200#pg10> Theorem 2 shows that heterogeneous information ages occur infinitely often almost surely under independent, nondegenerate stochastic information arrival <ref:2610.11200#pg10> Consequently, information divergence—and hence asymmetric information—recurs infinitely often almost surely <ref:2610.11200#pg10> Endogenous Updating Dynamics Agent i chooses an updating intensity ui ∈ (0, 1] <ref:2610.11200#pg10> The resulting information-age process satisfies Ai(t + 1) = Z i(t + 1) = 1, A i(t) + 1, Z i(t + 1) = 0 <ref:2610.11200#pg10> The resulting information-age Markov chain has stationary distribution Pr(Ai = a ui) = u i(1 − u i) a, a = 0, 1, 2, 3...

Ruth: <ref:2610.11200#pg10> Optimal Updating Intensity Agent i chooses ui to minimize Ji(ui) = lambda i L i(ui) + Ci(ui) <ref:2610.11200#pg10> Proposition 2 implies that perfectly current information need not be optimal, and the optimal intensity satisfies ui∗ Corollary 1 shows that identical agents choosing the same interior updating intensity u star generate heterogeneous information ages, and hence information divergence, recur infinitely often almost surely <ref:2610.11200#pg10> Threshold Equilibria and Certification Proposition 5 characterizes threshold equilibria in which state dynamics determine valuation differences across types <ref:2610.11200#pg9> Under the stationary AR(1) process with geometric information ages generated by updating probability 0 < u < 1, a threshold m is an equilibrium cutoff if and only if kL(m; u) ≤ k ≤ kH(m; u) <ref:2610.11200#pg10> Corollary 3 demonstrates that more frequent information arrival raises the valuation of the silent pool and weakly lowers both the smallest and largest equilibrium cutoffs wherever positive-certification equilibria exist <ref:2610.11200#pg10> Numerical Illustration Table 2 reports the equilibrium cutoffs under geometric information ages for three updating intensities <ref:2610.11200#pg10> Figure 3 illustrates the results, showing that more frequent updating lowers the cutoff, as Corollary 3(iii) predicts, from 4 at u = 0.10 to 2 or 3 at u = 0.50 <ref:2610.

Title and authors: Ruth: So we're talking about this paper now called Information Divergence and Endogenous Evolution of Asymmetric Information. It seems to be looking at how agents can get different predictions even if they are all rational and observing the exact same underlying situation at different times.

Warren: Right, Ruth. The authors, Ratna K Shrestha, Gaurav Subedi, and Ayush Shrestha from the Sauder School of Business at UBC, they set up this framework to show that standard theories about information asymmetry—like private signals or costly acquisition—don't have to be the only way we explain why people disagree.

Iris: The core idea here is that even when agents are identical and see a common Markov state, their predictive distributions can diverge just because their information ages are different. That's the mechanism they focus on first.

Ruth: Exactly. It moves the source of asymmetry away from private hidden things and focuses it purely on when you observe something. They introduce the concept of information age, which is simply how much time has passed since an agent last saw a state in their observation record.

Warren: And they show that when this information age is involved, if the way the system transitions between states separates these ages—if observations from different times lead to different future predictions—then independent arrivals of new observations can cause these differences to happen over and over again almost surely, even if everyone has the exact same probability of observing something at any given moment.

Iris: That's what they call information divergence, and it’s the key driver for this whole asymmetry problem. They make this divergence a recurring event under certain conditions regarding how observations arrive.

Ruth: But they don't just stop there with the recurrence; they look at how agents actually behave when they have to choose how often to update their information. That's where things get endogenous.

Warren: The paper says agents can choose an updating intensity, let's call it u, which is the probability that they decide to observe something at a given step. When this updating process is costly, agents have to weigh the expense of refreshing their information against the risk of acting on old information.

Iris: And when you look at how that updating process works mathematically in this model, it sets up a Markov chain for those information ages, and they find a stationary distribution that depends on that chosen updating intensity u.

Title and authors: Ruth: So they found that the optimal updating intensity isn't always about being perfectly current. It turns out there's an optimal intensity u star where agents can still generate these heterogeneous information ages, which means divergence keeps happening over and over almost surely, even if everyone chooses the same updating strategy.

Warren: That’s interesting for trading strategies because it suggests that running the exact same observation schedule doesn't guarantee identical information structures among your peers. It opens up a whole new way for agents to create informational differences internally without relying on hidden private data or special acquisition costs.

Iris: And when you look at the consequences in a stationary Gaussian AR one environment, they find that the outcome depends on two things: the common staleness and the age gap between agents. Holding that age gap fixed, increasing how stale everyone is raises each agent's uncertainty but it actually reduces how much their individual forecasts disagree with each other.

Ruth: That sounds counterintuitive for a moment—how can disagreement go down when uncertainty goes up? It leads to what they call stale consensus, where you have low disagreement even if the overall uncertainty about the future state is high.

Warren: It means that agreement between agents in their forecasts might not be a sign that everyone has perfectly updated information or better knowledge than others. Instead, it could just be that both of them are relying on observations that are sufficiently old at the same time.

Iris: That ties directly into the incentive structure for revealing information age. They find that as common staleness increases, the value of certifying how fresh your data is goes down because both sets of observations are becoming stale simultaneously.

Ruth: This leads to some interesting threshold results when you have multiple levels of information ages. Instead of everyone telling everyone exactly how old their data is, there’s a point where sufficiently fresh agents decide to certify their exact age while the older ones just pool together.

Warren: And they characterize this with threshold equilibria based on parameters like k lambda and the functions L and H. It’s a mathematical way of saying that there's a cutoff for when certification becomes economically worthwhile versus when it’s better to just stick with a simpler pooling strategy.

Iris: When you look specifically at geometric information ages, they find that more frequent updating doesn't necessarily help everyone uniformly. Instead, increasing the frequency of updates weakly lowers both the smallest and largest equilibrium cutoffs wherever those certification equilibria exist.

Title and authors: Ruth: So to wrap up this section, the paper suggests that this whole stale consensus effect explains why forecast disagreement and uncertainty can move in opposite directions depending on how old your underlying information is.

Warren: It points toward a new way to look at agreement across different institutions—that agreement might reflect common staleness rather than everyone having the most precise current information.

Iris: And the practical implication is that this framework bears on policy stuff, like things such as open banking and standardized observation timestamps, which could potentially reduce those age-based asymmetries and lower those certification costs we talked about.

Ruth: Exactly. It cautions us against treating agreement between institutions as a perfect signal of accurate assessments, suggesting that we need to be careful about what that agreement actually means in terms of information freshness.

Warren: Before we wrap up this discussion on the Information Divergence and Endogenous Evolution of Asymmetric Information paper, Iris has some final thoughts on what this whole mechanism implies for testing these kinds of models.

Iris: I think it’s important to remember that the model identifies how information age can decouple disagreement and uncertainty, which is a structural feature we need to test in real-world data where we can measure those age gaps directly.

Ruth: And I think Warren has a point about the practical implications for AI systems that try to forecast; they need to design their updating strategies with this staleness effect in mind if they want to get better results.

Warren: Yeah, and it’s important for us as traders or those building these models to realize that perfect timeliness isn't always the best bet if the cost of keeping up is too high.

Iris: So we’ve seen how information divergence recurs under independent arrivals, how updating intensity drives the age distribution, and how common staleness creates this strange consensus effect in AR one environments.

Ruth: And it really highlights that when you look at predictive distributions over a common state, the time since your last observation is as crucial as what the last observation actually was.

Warren: So for anyone listening who’s thinking about multi-agent forecasting or financial modeling, this paper gives you a reason to question whether agreement is always a good sign of quality.

Iris: That's the main structural contribution here, showing how endogenous updating dynamics can generate persistent informational differences that don't come from private signals.

Ruth: And we’ve got the full picture now on this paper, Information Divergence and Endogenous Evolution of Asymmetric Information.

The paper's summary: Ruth: So, to recap what we just covered, this paper shows how agents can end up making different predictions even when they are all rational and looking at the exact same underlying situation at different times.

Warren: Right. It’s not about some secret private signal or someone spending a lot of money to get better data. It comes down purely to timing.

Iris: Exactly, because the paper identifies "information age" as this key thing—it's just how much time has passed since an agent last saw a state in their observation record.

Ruth: And that age, combined with the most recent state they saw, is what actually determines their predictive distribution over what’s coming next.

Warren: The authors find that if the way the system moves between states separates these ages—meaning observations from different times lead to different future predictions—then independent new observations cause these differences to happen again and again almost surely.

Iris: That recurrence they show is called "information divergence," and it’s what creates that asymmetry we're talking about.

Ruth: But the paper gets more interesting because it looks at how agents actually make choices about how often they update their information, which is endogenous.

Warren: They let agents choose an updating intensity, which is just the probability they decide to observe something at any given step. And when that process costs money, the AI has to balance that cost against the risk of acting on old information.

Iris: The mathematical setup for that updating process shows a Markov chain for those information ages, and they find a stationary distribution based on how frequently you choose to update.

Ruth: The paper then says there’s an optimal updating intensity, which means perfectly current information isn't always the best strategy if the cost of refreshing is too high.

Warren: They even show that identical agents choosing the same update schedule can still create these different information ages over and over again, which fuels that divergence.

Iris: And when you look at how this plays out in a steady state, they find that common staleness and the gap between ages matter more than just having perfect information.

Ruth: They call this "stale consensus," where even if everyone disagrees on what the future state is, their forecasts might actually be very similar because they are all relying on observations that are old at the same time.

Warren: So instead of agreement being a sign that one person has better data, it could just mean both people are operating under a shared condition of staleness.

Iris: This leads to how agents decide whether to certify their freshness—if they should tell everyone how old their data is—and they find threshold equilibria based on those age levels.

Ruth: And for those who only listen, this means we can't just assume consensus means accuracy in a system where observation timing matters so much.

Warren: It’s a big shift because it suggests that the way information gets into an agent’s hands, not just what the information itself is, drives the divergence.

Iris: The next part they look at is how this plays out specifically in systems where states move like a random walk.

Ruth: They find that in those random walk environments, if you don't update often enough, your uncertainty about the actual state you’re predicting grows linearly without any bound.

Warren: That means the more old your information gets in that specific context, the bigger your prediction error gets over time.

Iris: The authors use this to argue that disagreement can actually understate how much uncertainty there really is precisely when the underlying information is becoming stale.

Ruth: This whole mechanism gives us a reason why forecast disagreement and uncertainty can move in opposite directions.

Warren: It also points toward policy ideas, like standardized timestamps or open banking, as ways to reduce those age-based asymmetries across different institutions.

Iris: And it’s a warning against treating agreement between groups as proof that everyone has made better assessments than the others.

Ruth: So the paper sets up this structural reason why disagreement and uncertainty can move in opposite directions based on how old your data is.

Warren: Before we look at what this means for actual trading, Iris wants to check if this identification holds up when we change the rules around observation arrival.

The paper's improvements: Ruth: So we just talked about how information divergence comes from observation timing, and now we’re looking at what the authors suggest to make this framework even more useful.

Warren: They basically say they could use AI to model how these agents actually develop these different information structures, which is a huge step forward for simulation.

Iris: That's right. They propose that AI systems can design their own updating schedules, choosing an intensity, which we called u in the paper earlier.

Ruth: And this allows the AI to learn that being perfectly current isn't always the best move if the cost of refreshing is too high.

Warren: Exactly. The paper suggests this makes sense for real-world trading where keeping up with every tiny tick is too expensive, so an AI might rationally accept a small chance of acting on stale data.

Iris: That’s interesting because it moves the model from just describing agents to suggesting how an agent would actually behave when they are trying to minimize a cost function.

Ruth: And then there’s the concept of "stale consensus," which isn't just a description, but a tool for interpretation.

Warren: The authors want us to use this effect to distinguish between genuine informational precision and just common staleness in market data.

Iris: They point out that when you see low forecast dispersion, you need to check if that’s because everyone is actually right about the current state or if they’re all just relying on old information.

Ruth: And they suggest a way for AI to figure out the optimal threshold for certifying information age based on those staleness levels.

Warren: They basically want AI to find that cutoff point, let's call it m star, where you have one group of agents who certify their freshness and another group that just pools their old data.

Iris: That’s a bit different from the math we used to find the cutoffs, because now the system is deciding on certification based on those staleness conditions rather than just a fixed rule.

Ruth: It seems like they are trying to make the model more flexible so it can explain why two institutions might agree even if their data freshness is wildly different.

Warren: It’s about building a simulation where the agents have an incentive structure that reflects real-world constraints on how much time they can afford to wait for new information.

Iris: And the geometric information ages part, where more frequent updates lower those cutoffs, provides a quantitative rule for how updating frequency affects which information ages are actually worth certifying.

Ruth: So to summarize these improvements, we’re moving from just finding the patterns to designing AI agents that can choose their own update rules based on cost and staleness thresholds.

Warren: It shifts the focus from just "what is the information?" to "how do agents use the time they have?"

Iris: This opens up a new way to test if these kinds of age-based asymmetries really exist in real datasets, as we can now look for that staleness signature.

Conclusion: Ruth: So we’re wrapping up on "Information Divergence and Endogenous Evolution of Asymmetric Information." We’ve seen how agents can create different predictions based purely on when they observe information, not just what it is.

Warren: It really shows that the structure of observation itself—the timing—can be a massive source of hidden asymmetry in markets or any multi-agent system.

Iris: The authors show that this divergence recurs almost surely under independent arrival processes, meaning we don't have to worry about it being a one-off glitch.

Ruth: And they’ve given us tools, like the optimal updating intensity, to figure out how agents make these choices when there’s a cost involved in refreshing their data.

Warren: It suggests that if you want to run a strategy where agents are behaving predictably, you have to account for this endogenous evolution of information ages.

Iris: The whole point is that we need to look at agreement across institutions not as proof of accuracy, but as a potential sign of common staleness.

Ruth: That’s the big takeaway for me—we can’t always assume consensus means everyone has current, perfect information.

Warren: It changes how we view market signals; sometimes the convergence you see is just everyone relying on old data simultaneously.

Iris: And it gives us a way to test if this age-based asymmetry is actually present in the real world by looking for that specific divergence pattern in our observation records.

Ruth: So, this paper provides a structural reason why disagreement and uncertainty can go in opposite directions depending on how old the underlying information is.

Warren: It’s a framework for understanding how temporal dynamics create persistent informational differences among identical rational agents.

Iris: It really highlights that the way information gets into an agent's hands, not just what the information itself is, drives the divergence.

Ruth: We’ve seen how they use this model to suggest policy ideas like standardized timestamps to reduce those age-based asymmetries across different institutions.

Warren: That points toward a more robust way of building systems that doesn't rely on perfect real-time data feeds from everyone.

Iris: It also gives us a testable implication, which is that if we hold the forecast dispersion fixed, the errors should be larger when the information underlying those forecasts is older on average.

Ruth: So, to conclude this paper, "Information Divergence and Endogenous Evolution of Asymmetric Information" suggests that timing is just as important as content in understanding agent behavior.

Warren: It’s a solid piece for anyone studying multi-agent forecasting or even complex financial modeling where information flow is messy.

Iris: We're looking forward to seeing how other researchers apply these staleness concepts to different state spaces next.

Sauder School of Business, University of British Columbia · RALP Technologies · Global Relay Communications

econ.TH

Submitted: 2026-10-08

Updated: 2026-11-08

Comments: 30 pages

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 92/100

The gist: Standard theories trace information asymmetry to privately held information, heterogeneous signals, or costly information acquisition <ref:2610.11200#pg6> The paper identifies a distinct mechanism…

Key concepts

Information Age
This is the time elapsed since an agent last observed the state of the system. It is a key factor in determining an agent's predictive distribution. The paper shows that different information ages can lead to different predictions, even when agents are otherwise identical.
Age-Separating Transition Kernel
This occurs when the rules governing how the state evolves depend on the age of the observation. If this condition holds, observations from different time points (different ages) will naturally generate distinct predictive distributions for an agent.
Stale Consensus
This is a phenomenon where agents agree on forecasts even if their underlying information is old. This happens because agents might rely on sufficiently stale observations, leading to low disagreement despite high uncertainty about the current state.

Terminology

Summary

Standard theories trace information asymmetry to privately held information, heterogeneous signals, or costly information acquisition <ref:2610. p. 011200#pg6> The paper identifies a distinct mechanism: otherwise identical rational agents observing a common Markov state at asynchronous times can hold different predictive distributions because their information differs in age <ref:2610. p. 011200#pg10> When the transition kernel separates information ages, independent Bernoulli observation arrivals make such divergence recur infinitely often almost surely, even with identical arrival probabilities <ref:2610. p. 011200#pg10> When agents instead choose costly updating intensities, independent update realizations can regenerate heterogeneous information ages among ex ante identical agents <ref:2610. p. 011200#pg10>

Mechanism of Information Divergence

The core mechanism identified is information age, defined as the time elapsed since an agent's most recent observation <ref:2610.11200#pg9> Together, the last observed state and its age determine the agent’s predictive distribution over the current state <ref:2610. p. 111200#pg9> When transition kernel is age-separating—so observations of different ages generate distinct predictive distributions—asynchronous observation produces what we call information divergence <ref:2610. p. 111200#pg10> Under independent, nondegenerate Bernoulli arrivals, this divergence recurs infinitely often almost surely, even when all agents face the same arrival probability <ref:2610. p. 111200#pg10>

Endogenous Updating and Staleness

The framework endogenizes information updating by allowing agents to choose the intensity of a stochastic observation process <ref:2610.11200#pg10> When updating is costly, agents trade off the resource cost of maintaining a higher refresh rate against the expected loss from acting on stale information <ref:2610. p. 111200#pg10> In a stationary Gaussian AR(1) environment, the consequences depend not only on common staleness as well as the age gap <ref:2610. p. 011200#pg9> Holding the age gap fixed, greater common staleness raises each agent's predictive uncertainty but reduces the disagreement between their predictive forecasts <ref:2610.11200#pg9> This generates stale consensus: low disagreement despite high uncertainty <ref:2610. p. 011200#pg9>

Consequences of Common Staleness

Common staleness also shapes incentives to disclose information age, as the value of certifying relative freshness declines as both observations become stale <ref:2610.11200#pg10> With multiple ordered information ages, costly certification yields a threshold equilibrium in which sufficiently fresh agents certify their exact age while sufficiently stale agents pool <ref:2610. p. 211200#pg10> Under geometric information ages, more frequent updating weakly lowers the smallest and largest equilibrium disclosure cutoffs wherever positive-certification equilibria exist <ref:2610. p. 011200#pg10> This stale consensus effect explains why agreement need not indicate better information: agents may issue similar forecasts because both have current information or because both rely on sufficiently stale observations <ref:2610. p. 211200#pg9>

Linear-Gaussian Specialization and Economic Implications

In the linear Gaussian AR(1) specialization, predictive uncertainty depends on three factors: environmental volatility (sigma 2), state persistence (rho), and information age (Ai(t)) <ref:2610.11200#pg10> For a given sigma squared and rho! p. 8= 0, predictive uncertainty increases with Ai(t) as unobserved innovations accumulate <ref:2610. p. 111200#pg10> Under the random walk process, predictive uncertainty about the current payoffrelevant state increases linearly with Ai(t), without bound <ref:2610. p. 611200#pg10> This stale-consensus effect implies that disagreement can understate uncertainty precisely when the underlying information is old <ref:2610. p. 611200#pg9>

Applications and Conclusions

The model provides a structural reason why forecast disagreement and uncertainty can move in opposite directions <ref:2610.11200#pg9> This stale-consensus result implies that agreement across institutions may reflect common staleness rather than precise current information <ref:2610. p. 611200#pg10> The framework also bears on information-sharing policy: open banking and standardized observation timestamps could reduce age-based asymmetries and certification costs <ref:2610. p. 011200#pg10> The staleconsensus result likewise cautions against treating agreement across institutions as evidence of accurate assessments <ref:2610. p. 211200#pg9> The model also yields a testable implication: holding observed forecast dispersion fixed, forecast errors should be larger when the information underlying those forecasts is older on average <ref:2610. p. 011200#pg9>

--- Page 3

Information divergence occurs at date t if there exist agents i ≠ j such that their predictive distributions over the current state differ: mu i(t) ≡ Pr(X t ∈ ⋅∣ ∣ I i(t)) ≠ Pr (X t ∈ ⋅∣ ∣ I j(t)) ≡ mu j(t) <ref:2610.11200#pg9> Different information ages need not, in general, imply different predictive distributions; Definition 4 states the condition under which they do <ref:2610. p. 811200#pg9>

Age-Separating Transition Kernel

The transition kernel Q is age-separating if for any a ≠ a' and x, x' ∈ X, Q a(∙∣ x) ≠ Q a'(∙∣ x') <ref:2610.11200#pg9> Thus observations of different ages induce distinct predictive distributions, regardless of the realized values of the most recent observations <ref:2610. p. 911200#pg9>

Dynamic Emergence and Persistence

Theorem 1 isolates our central mechanism: asynchronous observation of an evolving payoffrelevant state can generate asymmetric information among otherwise identical rational agents <ref:2610.11200#pg10> Theorem 2 shows that heterogeneous information ages occur infinitely often almost surely under independent, nondegenerate stochastic information arrival <ref:2610. p. 111200#pg10> Consequently, information divergence—and hence asymmetric information—recurs infinitely often almost surely <ref:2610. p. 911200#pg10>

Endogenous Updating Dynamics

Agent i chooses an updating intensity ui ∈ (0, 1] <ref:2610.11200#pg10> The resulting information-age process satisfies Ai(t + 1) = Z i(t + 1) = 1, A i(t) + 1, Z i(t + 1) = 0 <ref:2610. p. 411200#pg10> The resulting information-age Markov chain has stationary distribution Pr(Ai = a ui) = u i(1 − u i) a, a = 0, 1, 2, 3. p. 8.. <ref:2610.11200#pg10>

Optimal Updating Intensity

Agent i chooses ui to minimize Ji(ui) = lambda i L i(ui) + Ci(ui) <ref:2610.11200#pg10> Proposition 2 implies that perfectly current information need not be optimal, and the optimal intensity satisfies ui∗ Corollary 1 shows that identical agents choosing the same interior updating intensity u star generate heterogeneous information ages, and hence information divergence, recur infinitely often almost surely <ref:2610. p. 011200#pg10>

Threshold Equilibria and Certification

Proposition 5 characterizes threshold equilibria in which state dynamics determine valuation differences across types <ref:2610.11200#pg9> Under the stationary AR(1) process with geometric information ages generated by updating probability 0 < u < 1, a threshold m is an equilibrium cutoff if and only if kL(m; u) ≤ k ≤ kH(m; u) <ref:2610. p. 211200#pg10> Corollary 3 demonstrates that more frequent information arrival raises the valuation of the silent pool and weakly lowers both the smallest and largest equilibrium cutoffs wherever positive-certification equilibria exist <ref:2610. p. 011200#pg10>

Numerical Illustration

Table 2 reports the equilibrium cutoffs under geometric information ages for three updating intensities <ref:2610.11200#pg10> Figure 3 illustrates the results, showing that more frequent updating lowers the cutoff, as Corollary 3(iii) predicts, from 4 at u = 0. p. 010 to 2 or 3 at u = 0.50 <ref:2610.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Endogenous Information Divergence Modeling: AI systems can be designed to model how otherwise identical rational agents become differently informed when they observe the same evolving payoff-relevant state at different times. This allows for simulating scenarios where information asymmetry arises solely from observation timing, rather than private signals or costly acquisition.

  2. Optimal Information Updating Strategy: AI agents can choose an updating intensity u to minimize a cost function J i (u i) = lambda i L i (u i) + C i (u i). This means the AI can learn that perfectly current information need not be optimal and rationally accept a positive probability of operating with stale information if the cost of updating is high.

  3. Stale Consensus Interpretation: The system can distinguish between genuine informational precision and common staleness by observing that low forecast dispersion need not indicate high information quality. This enables the AI to recognize when agreement among agents reflects common staleness rather than precise current information, which is crucial for interpreting consensus in financial forecasting or multi-agent AI.

  4. Age-Based Disclosure Thresholds: An AI can determine the optimal cutoff m star for certification strategies by solving the condition where "k <= V(m star) - V̄(m star). This allows the system to implement threshold equilibria, such as deciding when fresher types certify and sufficiently stale types pool," in contexts like credit or insurance.

  5. Geometric Age Thresholds: For systems with geometric information ages, an AI can identify that more frequent information arrival raises the valuation of the silent pool and weakly lowers both the smallest and largest equilibrium cutoffs. This provides a quantitative rule for how updating frequency affects which information ages are worth certifying.

Sources

Related papers