Traffic-Grounded Synthetic Buyers

• Traffic-grounded synthetic buyers are empirically constructed agent populations that mimic observed transactional behaviors from urban mobility and e-commerce environments.
• Their construction involves detailed calibration of behavioral archetypes and session profiles, ensuring realistic decision policies and simulation fidelity.
• Applications include testing interventions such as tradable credit schemes and online A/B experiments, validated against empirical data for predictive accuracy.

Traffic-grounded synthetic buyers are empirically constructed agent populations whose behaviors, states, and success criteria are calibrated against observed traffic from real human users in transactional environments such as transportation networks and e-commerce storefronts. These agents are designed to ground their decision policies, heterogeneity, and transactional logic in realistic behavioral distributions extracted from historical interaction data, enabling rigorous and policy-relevant testing of system interventions such as tradable credit schemes (TCS) or ecommerce user interface (UI) modifications. Two principal domains exemplify traffic-grounded synthetic buyers: activity-based traveler agents in urban mobility simulations for TCS evaluation (Liu et al., 17 Feb 2025), and LLM-based buyer agents simulating online shopping cohorts for A/B testing (Castelo et al., 1 Feb 2026).

1. Foundational Principles

Traffic-grounded synthetic buyers leverage empirical distributions of user sessions, behavioral clusters, and intent states to generate agent populations whose micro-decisions reflect the diversity and statistics of real observed traffic. This grounding operates on two intertwined levels:

• Behavioral Archetype Assignment: Agents are mapped to empirically discovered cohorts or clusters, each representing a distinct behavioral regime (e.g., highly-engaged buyers, low-exploration browsers).
• Session Profile and Intent Calibration: Conditional on cohort assignment, an agent's decision policies—such as travel utility calculation, product search depth, purchase-readiness, or price sensitivity—are drawn from empirical session features or conditioned intent statistics.

The empirical grounding enables subsequent simulations, interventions, or policy experiments to reflect counterfactual impacts at the micro (agent) and macro (system) levels with fidelity to the distributional properties of actual user traffic.

2. Construction Methodologies

The instantiation of traffic-grounded synthetic buyers differs by domain, but universally involves careful synthesis of agent attributes, state transitions, and intent logic.

Urban Mobility/TCS Simulators

A canonical workflow, as implemented in SimMobility (Liu et al., 17 Feb 2025), includes:

• Population and Activity Initialization: Generate $N$ synthetic travelers using census or survey data, assigning home/work locations, tour structures, and alternative path/route sets. Attributes such as value of time $VOT_n$, schedule-delay penalties $(SDE_n, SDL_n)$, and choice sets are drawn from empirically calibrated distributions.
• Credit Account and Budget Dynamics: Each agent manages a credit account $x_n^d(t)$ governed by allocation rates, budget constraints, and toll costs, with explicit logic for buying/selling credits based on real-time needs and market prices (see formulas below).
• Utility-based Choice Modeling: Travel alternatives are evaluated using a utility function incorporating agent-specific time, schedule, and credit cost terms, with the credit price $p_d$ dynamically updated via market clearing.

E-Commerce Synthetic Buyer Agents

SimGym constructs agent cohorts by:

• Session-level Feature Engineering: Extract $d$-dimensional vectors $x_s$ from production clickstream logs, capturing engagement, exploration, funnel progression, and economic features.
• Clustering: Run $k$-means (typically $k = 5$) on standardized session features, assigning each session to a cohort $c(s)$.
• Cohort Priors and Intent Distribution: The traffic-grounded prior over cohorts is $P(c) = N_c / N$. Within each cluster, purchase vs. browse intent distributions approximate empirical add-to-cart (A2C) rates $a_c$, ensuring realistic representation of buyer readiness and category interest.

3. Agent Behavior and Interaction

Mobility Credit Market Agents

Credit-budgeted decisions are central, with market-structured buyer logic as follows:

• Credit Account Dynamics:

For trip $i$ at $t = t^{dep}$:

$$x_n(t+\Delta t) = \begin{cases} \min[x_n(t) - g(t) \cdot Dist_i + r \cdot \Delta t, l \cdot r] & \text{if } x_n(t) \geq g(t) \cdot Dist_i \ r \cdot \Delta t & \text{otherwise (must buy shortfall)} \end{cases}$$

(other rules handle selling, idling, constraints as in (Liu et al., 17 Feb 2025) eqns. (1)-(4)).

• Credit Market Transactions:

$B(y) = y \cdot p_d \cdot (1+\hat{f}_b) + f_b$

$S(y) = y \cdot p_d \cdot (1-\hat{f}_s) - f_s$

(eqns. (6)-(7)), with market prices updated per

$p_{d+1} = \max[p_d + k \cdot Z_d, 0]$

where $Z_d$ is the net credit demand (eqn. (5)).

E-Commerce SimGym Agents

• Browser Environment: Each agent operates in a live browser, receiving observations from the accessibility tree, with persona and intent downloaded from data-discovered cohort distributions.
• LLM-based Planning: The LLM receives a structured prompt containing intent, persona, action history, current page state, and guardrails, outputting JSON-encoded action plans.
• Policy Execution: Possible actions include element clicks, scrolling, text entry, and navigation. Session memory and step-by-step updates maintain episode coherence.
• No Learning Loop: In SimGym's baseline, the LLM acts as a zero-shot planner without explicit reward or RL-based fine-tuning. Fidelity is measured ex post against outcome alignment statistics (Castelo et al., 1 Feb 2026).

4. Simulation, Calibration, and Experimentation

Mobility Simulation (TCS)

• Parameterization: Key simulation parameters—allocation rate $r$, lifetime $l$, toll function $g(t)$, transaction fees—are calibrated from empirical or policy targets.
• Optimization Loop: Toll and allocation parameters are optimized via Bayesian Optimization (BO) using a Matérn kernel GP, with social welfare $SW(x)$ as the objective:

$$SW(x) = \sum_n [U_n(x) - U_{n,base}] / |\beta_{cost,n}|$$

The loop iterates, running the agent-based simulator for $x_{next} = \arg\max_x UCB(x)$, until convergence.

• Performance Metrics: Outcomes are assessed via day-to-day convergence (inconsistency metric), realized utility trajectories, network-level quantities (flow, accumulation, travel-time indices), market equilibrium statistics (credit price, trading volumes), and heterogeneity/equity effects (variation by $VOT$ quantile).

E-Commerce SimGym

• Session Generation: For each test, a weighted sample of $N$ agents per cohort is instantiated ($N \approx 600$), mirroring live cohort distributions.
• A/B Simulation: Each agent plays both control and treatment storefronts with identical seeds and persona. Session-level metrics (A2C, revenue) are collected.
• Cohort-weighted Aggregation: Treatment lift estimates aggregate cohort-weighted outcome differences:

$$\hat{\Delta} = \sum_{c=1}^k w_c \left(\mathbb{E}_{s\sim S_c^{treat}} Y(s) - \mathbb{E}_{s\sim S_c^{ctrl}} Y(s) \right)$$

with metrics computed for conversion and revenue per session.

• Bootstrap and Cycle-time: Alignment and correlation metrics are bootstrapped to guide sample size selection; N=600 is chosen for empirical stability. Full A/B simulations execute in under one hour, contrasting with weeks required for live experiments.

5. Evaluation and Validation

Predictive Validity

• Ground-truth Construction: Changes are validated only when external confounders (promotions, assortment, seasonalities) are absent, confirmed by double/debiased machine learning.
• Alignment Metrics:
  • Directional alignment rate: Fraction of shops where simulated and human treatment effects align in sign.
  • Bayesian sign alignment probability: Posterior probability that simulated and human effects agree in direction.
  • Pearson correlation ($r$): Between agent and human treatment lifts.
• Findings: Even without alignment post training, SimGym agents achieve state-of-the-art outcome alignment with observed human behavior.

TCS Simulator Outcomes

Day-to-day consistency, equilibrium credit prices, network flow, marginal gains, and heterogeneity in welfare improvements are computed as validation metrics. Bayesian Optimization converges to welfare-optimal designs; agent-based simulation matches known theoretical properties of TCS under behavioral assumptions (Liu et al., 17 Feb 2025).

6. Limitations and Future Directions

• E-Commerce Agents: Limiting observations to accessibility trees excludes sensitivity to purely visual changes; broader UI interventions remain untested. Absence of RL or supervised fine-tuning may reduce behavioral nuance. Future prospects include neural persona extraction and automated closed-loop UI optimization (Castelo et al., 1 Feb 2026).
• Mobility TCS Agents: Behavioral assumptions—for example, heuristic-based credit selling—may influence outcomes. Extensions could incorporate mode-choice models, network assignment, or richer traveler heterogeneity.

A plausible implication is that the traffic-grounded synthetic buyer paradigm enables robust, interpretable, and empirically validated offline testing of policies and system changes, accelerating experimentation while reducing exposure of live users or travelers.

---

Domain	Empirical Grounding	Agent Architecture
Urban Mobility (TCS)	Census/survey travel data	Activity-based agent models, credit/market actions
E-Commerce (SimGym)	Clickstream session data	LLM browser agents w/ cohort persona, prompt-based planning

Both frameworks demonstrate that detailed session-level calibration is essential for replicating real-world traffic responses in agent-based simulations. These methodologies underpin both theoretical and practical advances in the design, evaluation, and deployment of interventions in dynamic market and infrastructure systems.