Agent-Based Asset Pricing Model

• Agent-Based Asset Pricing Model is a computational framework where heterogeneous agents interact, revealing authentic asset price dynamics.
• It employs deep reinforcement learning and a discrete limit order book to simulate market microstructure and capture stylized facts like fat tails and volatility clustering.
• The framework uses optimal transport for calibrating agent traits, leading to endogenous emergence of differentiated trading strategies and macro-scale market properties.

An agent-based asset pricing model (ABAPM) is a computational financial market framework in which the price dynamics of assets arise endogenously from the interactions of a population of explicitly modeled agents, each with specified behavioral, informational, and preference heterogeneity. Unlike representative-agent or Walrasian paradigms, ABAPMs accommodate bounded rationality, learning, behavioral biases, heterogeneous risk preferences, and complex microstructure, providing rich explanatory power for observed “stylized facts” of asset returns such as fat tails, volatility clustering, and regime shifts. Below, the organizing features, methodologies, calibration pipelines, and emergent phenomena of modern ABAPMs are detailed, focusing on models where the interaction of agent learning and preference heterogeneity is central to the collective market dynamics (Hashimoto et al., 7 Nov 2025).

1. Model Environment and Price Formation

Most contemporary ABAPMs simulate a central electronic limit order book (LOB) in discrete event time. A finite population of agents (typically $n=50$–$200$) trade a single risky asset (the stock) against a risk-free asset (cash), submitting limit orders $(v_t^j, p_t^j)$ where $v_t^j$ denotes signed volume and $p_t^j$ is the limit price. The LOB matches orders by price-time priority; transaction prices set the evolving mid-price

$$p_t^{mid} = \frac{1}{2}(\text{best ask} + \text{best bid}),$$

with no closed-form price impact assumed beyond the microstructure-induced nonlinearities.

Agents update portfolios (risky position $w_t$, cash $c_t$), observing a state vector containing both aggregated market microstructure variables (returns, volatility, order book depths) and their individual portfolio state and trait parameters. Each step, one agent is uniformly selected to submit, introducing asynchronous and stochastic market participation (Hashimoto et al., 7 Nov 2025, Wagner et al., 2014).

2. Agent Heterogeneity: Preferences, States, and Utility

A defining feature is the explicit parametrization of agent heterogeneity. In the recent "Emergence from Emergence" framework (Hashimoto et al., 7 Nov 2025), agents are assigned at the start of each rollout three independent traits: $$\sigma^j \sim \mathcal{N}(\mu^\sigma,(\lambda^\sigma)^2),\quad \alpha^j \sim \mathcal{N}(\mu^\alpha,(\lambda^\alpha)^2),\quad \gamma^j \sim \mathrm{Uniform}(\lambda^\gamma,\,\gamma^{\max})$$ where

• $\sigma^j$: degree of uninformedness (noise in fundamental signal reception),
• $\alpha^j$: CARA risk aversion,
• $\gamma^j$: time discount factor.

Agents optimize a discounted sum of immediate utilities, where the core reward $g(o_t^j,a_t^j;j)$ incorporates:

• A bounded utility transformation (arctangent),
• Realized profit/loss since last action,
• Portfolio-dependent risk penalties (short positions, negative cash),
• Illiquidity surcharges (order book depth),
• Fundamental mispricing penalties.

The agent’s optimization problem is

$$\max_{\pi^j}\;\mathbb{E}\Bigl[\sum_{i=1}^{\iota_j}(\gamma^j)^i\,g(o_{t_i^j}^j,a_{t_i^j}^j;j)\Bigr],$$

where $\pi^j$ is the policy and $o_t^j$ includes both local and personal trait information.

3. Learning Protocol and Policy Architecture

Agents learn via deep reinforcement learning, operationalized using a shared-parameterized neural network policy $\pi_\theta(a|o)$. All agents share weights but condition on their static trait parameters, enabling differentiated behaviors despite unified architecture. State vectors $o_t^j \in \mathbb{R}^{11}$ encode:

• Market variables (returns, book imbalances, volatility)
• Personal state (portfolio, cash, total wealth)
• Trait inputs ($\sigma^j, \alpha^j, \gamma^j$).

Actions specify order volume and relative price aggressiveness. Mappings ensure translation from normalized action space to permissible market orders: $$v_t^j=\lceil v_{\max}\,\tilde v_t^j\rceil,\qquad p_t^j= p_t^{mid} - r_{\max}\,\mathrm{sign}(\tilde v_t^j)\,\tilde r_t^j\,p_t^{mid}.$$

Training follows batched Proximal Policy Optimization (PPO): agent trajectories are collected into per-agent buffers; when a buffer is full, advantage estimates are computed and actor (policy) and critic value networks are updated in stochastic gradient steps.

4. Calibration, Simulation Protocol, and Matching to Stylized Facts

The trait-distribution hyperparameters ($\lambda^\sigma, \lambda^\alpha, \lambda^\gamma$) are fit via optimal transport to minimize distance between empirical "point clouds" of synthetic and real market return sequences, tail events, and volatility paths, ensuring generated price series conform to established empirical stylized facts. Simulation episodes run for $T_{sim}\approx 10^5$ steps, agents are redrawn at each rollout to enforce broad generalization, and learning rates/hyperparameters are tuned for stability (Hashimoto et al., 7 Nov 2025).

Baseline comparisons include:

• ZI-agents (pure random orderflow): fail to match fat tails or volatility clustering;
• FCN-agents (fixed heuristics with heterogeneity): partial reproduction of heavy tails, less clustering;
• Adaptive chartist/fundamentalist models: improve but insufficient matching of empirical regularities;
• Full heterogeneous RL agents: match full fat-tailed distributions ($\hat\alpha \approx 3$), volatility clustering ($\text{Corr}(|r_t|,|r_{t+\tau}|) \sim \tau^{-\zeta}$, $\zeta\in(0,1)$), and correct volume-volatility correlation.

5. Emergence, Differentiation, and Macro-Scale Market Properties

The principal finding is "two-stage emergence":

1. At the micro level, joint learning and heterogeneous preferences drive agents to develop differentiated strategies matching their trait profiles (risk aversion, myopia, information quality), spontaneously creating a population of "fundamentalists," "momentum traders," and "noise traders"—without exogenous role assignment or hand-coded strategy classes.
2. At the macro level, the interactions among these differentiated micro-roles produce aggregate price series exhibiting all canonical stylized facts: heavy-tailed return distributions (empirical kurtosis $\gg 3$), power-law decay in return autocorrelations, positive volume-volatility association, and volatility clustering.(Hashimoto et al., 7 Nov 2025)

Crucially, ablation experiments show that removing either learning (using only hand-coded heterogeneity) or agent heterogeneity (using fully homogeneous learners) destroys the emergence of realistic market properties and collapses aggregate welfare.

6. Structural and Theoretical Extensions

The core paradigm in (Hashimoto et al., 7 Nov 2025) is a constructive framework for hierarchical emergence: embedding agent identities directly into both observation space and reward structure while training a shared policy to adapt to all variations simultaneously. This approach enables:

• Endogenous behavioral differentiation and dynamic "niche specialization" based on market conditions and evolutionary competitive feedback.
• A quantitative micro–macro bridge for the stylized facts: power-law returns, volatility clustering, and real trading regularities follow naturally from interactions—not from imposed distributional assumptions or ad hoc noise.

Coupling this RL-based heterogeneity paradigm with classical microstructure models (order-book matching, explicit liquidity, and endogenous order flow) as demonstrated yields agent-based asset pricing models with unprecedented empirical realism. The result is a unified platform for causal investigation into how the joint distribution of preferences and adaptive cognition generates both individual behavioral diversity and aggregate asset-pricing regularities (Hashimoto et al., 7 Nov 2025).