ShortageSim: Pharma Shortage Simulation

• ShortageSim is a multi-agent simulation framework that models pharmaceutical drug shortages by simulating strategic interactions among manufacturers, buyers, and regulatory agencies.
• It integrates a two-stage LLM-driven decision pipeline to capture bounded rationality and behavioral heterogeneity under conditions of information asymmetry.
• Empirical validations demonstrate improved shortage resolution timing through dynamic regulatory interventions and nuanced market response modeling.

ShortageSim is a multi-agent simulation framework for modeling and analyzing pharmaceutical drug shortages under conditions of information asymmetry. It leverages LLMs to emulate bounded-rational, strategic decision-making by manufacturers, institutional buyers, and regulatory agencies in response to supply disruptions and regulatory signals such as FDA shortage alerts. The framework advances simulation and intervention design by capturing real-world behavioral heterogeneity and opacity not addressed by traditional game-theoretic models.

1. Architecture and Framework Design

ShortageSim comprises four interconnected modules:

• Environment Module: Manages simulation state, inventory levels, demand shocks, and regulatory intervention logistics. Enforces information barriers between agents by controlling observable and hidden state variables.
• Agent System: Implements discrete agent roles—manufacturers, buyers (healthcare consortia), and the FDA—each with distinct decision processes. Every agent executes a two-stage LLM prompt pipeline:
  • Collector/Analyst: Synthesizes unstructured and partial input (including FDA communications and historical demand) to structured signals.
  • Decision Maker: Outputs investment, procurement, or announcement actions conditioned on synthesized context.
• Information Flow Component: Models asymmetric information—manufacturers have limited knowledge of competitors’ disruptions and investments; buyers access partial market signals.
• Simulation Controller: Orchestrates sequential production games across quarters, collecting decision traces for replay and analysis.

This design enables nuanced simulation of strategic interactions under uncertainty and regulatory influence, with agent-level reasoning fidelity contributed by the LLM backbone.

2. Simulation Game Dynamics

ShortageSim executes a sequential production game over multiple discrete periods (quarters):

• Disruption Modeling: Each period, manufacturers face probabilistic capacity disruptions; production capacity $c_i$ is reduced by a fraction $\delta$, with events tracked privately.
• Regulatory Announcements: Post-disruption, the FDA agent may issue either reactive alerts (in response to observed shortages) or proactive warnings (anticipatory), shifting market expectations.
• Agent Decision-Making: Manufacturers choose capacity investment and output; buyers adjust procurement and stockpile levels. Decisions are informed by individual beliefs and perceived regulatory signals.
• Demand Allocation and Market Clearing:
  • For disrupted manufacturers $i \in M_D$:
  • $q_i = \min\left\{\frac{D}{n},\ c_i\right\}$
  • (Equation 1: individual output is limited by either equal share of demand or capacity.)
  • Any unsatisfied demand from $M_D$ is reallocated among non-disrupted manufacturers $i \in M_U$:

  $$q_i = \min\left\{\frac{D}{n} + \frac{D_{\text{unfill}}}{|M_U|},\ c_i\right\}$$

  (Equation 2: incremented allocation for remaining suppliers.)

• Shortage Calculation: Aggregate supply is computed $Q = \sum_i q_i$; shortage level is $(D - Q)^+$.

This structure tightly integrates agent choices, hidden information, and the impact of regulatory signals on supply chain behavior.

3. Bounded-Rational Decision Model

Unlike classic models assuming Nash equilibrium and perfect rationality, ShortageSim explicitly models bounded rationality. For each agent:

• Perception Formation: Uses the Collector/Analyst LLM stage to infer market state from ambiguous or incomplete inputs.
• Decision Making under Uncertainty: Decisions are generated by the Decision Maker LLM prompt, incorporating heuristics, historical experience, regulatory cues, and risk factors, without assuming global optimality or full information.
• Behavioral Diversity: Agents may overreact, underreact, stockpile, or act conservatively according to local context and their interpretation of signals—aligning simulation behavior more closely to observed phenomena in pharmaceutical crises.

This allows ShortageSim to simulate heterogeneous, suboptimal, and context-sensitive responses, improving realism for policy testing.

4. Regulatory Signal Propagation and Impact

The FDA agent plays a pivotal role by issuing:

• Reactive Alerts: Triggered by materialized shortages (reported disruptions); tend to accelerate stockpiling and capacity responses.
• Proactive Warnings: Issued preemptively, aiming to dampen propagation of future disruptions via anticipatory behavior.

Announcement activity is quantified via the FDA Intervention Percentage (FIP):

$$FIP(\%) = \left(\frac{1}{T}\sum I_t\right)\times100$$

(Equation 3: $I_t$ is the announcement indicator for period $t$, $T$ total periods.)

Regulatory advisories alter risk perception, directly affecting agent policies (procurement, investment), and indirectly modifying the timing and resolution of shortage episodes.

5. Empirical Validation and Performance Metrics

ShortageSim is empirically validated against a dataset of 2,925 FDA shortage events and 51 resolved cases:

• Resolution-Lag Percentage (RLP): Measures simulation alignment to ground-truth shortage period durations:

$$RLP_j = 100 \times \frac{t_{\text{sim},j} - t_{\text{GT},j}}{t_{\text{GT},j}}$$

(Equation 4: $t_{\text{sim},j}$ is simulated resolution time, $t_{\text{GT},j}$ is ground-truth.)

• Results: In FDA-Disc (discontinued-disclosed) cases, ShortageSim reduced RLP by 83% over a GPT-4o zero-shot baseline (mean 1.4% vs. 8.42%). For cases without a disclosed cause (FDA-NR), it produced a conservative bias similar to baseline but reflected improved timing accuracy.
• Behavioral Analysis: Simulated trajectories demonstrate that regulatory announcements induce coordinated buyer stockpiling and manufacturer investment, especially under increased market opacity.

These outcomes substantiate the value of agent heterogeneity and bounded rationality for modeling real-world restoration timing in shortage scenarios.

6. Open-Source Dataset and Reproducibility

ShortageSim and the FDA shortage event dataset are publicly released at https://github.com/Lemutisme/Sortage_Management. The repository includes:

• Simulation code (multi-agent orchestration, LLM prompt design)
• Documentation detailing agent role templates, simulation parameters, and pipeline protocols
• Preprocessed event trajectories suitable for benchmarking policy interventions or alternative agent reasoning architectures

This enables broad reproducibility and extensibility for academic research on shortage management and regulatory strategies in information-scarce supply chains.

---

ShortageSim marks a methodological advance in the simulation of pharmaceutical shortages by integrating LLM-driven bounded-rational multi-agent modeling, strategic information asymmetry, and dynamic regulatory intervention. It achieves empirically validated improvements in supply restoration timing compared to zero-shot baselines and is built for open research into intervention effectiveness in complex supply chain networks (Cui et al., 1 Sep 2025).