Simulating Economic Shocks with AI

by

Author: Denis Avetisyan

Researchers are leveraging the power of large language models to generate realistic, data-driven stress tests for financial portfolios.

A scenario generation pipeline leverages large language models and information retrieval to produce structured economic forecasts-encompassing GDP, inflation, and interest rates-subject to both hard and soft plausibility constraints, then translates these forecasts into asset return projections via volatility scaling, linear principal component analysis, and a nonlinear polynomial factor incorporating textual amplification, with regime severity modulated by a mixing parameter $λ$ and performance benchmarked against deterministic baselines.
A scenario generation pipeline leverages large language models and information retrieval to produce structured economic forecasts-encompassing GDP, inflation, and interest rates-subject to both hard and soft plausibility constraints, then translates these forecasts into asset return projections via volatility scaling, linear principal component analysis, and a nonlinear polynomial factor incorporating textual amplification, with regime severity modulated by a mixing parameter $λ$ and performance benchmarked against deterministic baselines.

This paper presents a hybrid prompt-RAG pipeline for generating counterfactual macroeconomic scenarios and assessing portfolio risk.

Traditional macroeconomic stress testing struggles to efficiently incorporate diverse, unstructured data sources for realistic scenario generation. This is addressed in ‘LLM-Generated Counterfactual Stress Scenarios for Portfolio Risk Simulation via Hybrid Prompt-RAG Pipeline’, which details a novel pipeline leveraging large language models to create auditable, counterfactual macroeconomic scenarios for portfolio risk assessment. The study demonstrates that a hybrid prompt-retrieval approach yields coherent, country-specific stress narratives with stable tail-risk amplification, driven primarily by portfolio composition and prompt design. Could this framework offer a scalable and interpretable complement to existing stress-testing methodologies, ultimately enhancing financial risk management?

The Limits of Human Foresight in Financial Stress Testing

Historically, financial institutions have evaluated their resilience by subjecting balance sheets to predefined “stress tests” – simulations of adverse economic conditions. These tests, however, traditionally depend on scenarios painstakingly constructed by experts, a process inherently limited by human foresight and computational resources. The reliance on hand-crafted narratives means that tests often fail to adequately capture the breadth of potential shocks, particularly those stemming from novel or interconnected risks-like those seen in the 2008 financial crisis or more recently with geopolitical events. Because these scenarios are finite and pre-defined, they struggle to encompass the ‘unknown unknowns’ – the low-probability, high-impact events that pose the greatest threat to financial stability and can quickly render existing risk models obsolete. Consequently, institutions may underestimate their true vulnerability, creating a false sense of security and hindering effective risk management.

The creation of plausible and varied macroeconomic scenarios presents a significant hurdle in financial stress testing due to its intensive computational demands and reliance on specialized expertise. Simulating the complex interplay of economic factors – such as GDP, inflation, unemployment, and interest rates – requires sophisticated models and substantial processing power. Furthermore, defining the parameters and validating the realism of these scenarios necessitates deep economic understanding and judgment, often relying on the insights of seasoned financial analysts. This combination of computational cost and expert dependence creates a bottleneck, limiting the number of scenarios that can be effectively explored and hindering a truly comprehensive assessment of potential risks to financial institutions and the broader economic system. Consequently, institutions may struggle to adequately prepare for unforeseen shocks and emerging vulnerabilities.

The interwoven nature of modern financial systems and the accelerating pace of global change necessitate a shift beyond traditional, manually-constructed stress test scenarios. Contemporary economic shocks are rarely isolated; they propagate through complex networks via interconnected markets and institutions, demanding models capable of simulating a far broader range of potential crises. This growing complexity renders expert-driven scenario design both time-consuming and prone to overlooking systemic vulnerabilities. Consequently, automated approaches – leveraging machine learning, agent-based modeling, and high-performance computing – are becoming crucial for generating the diverse and realistic macroeconomic shocks required to comprehensively assess financial stability and anticipate emerging risks before they materialize. These methods offer the potential to explore a vastly expanded solution space, uncovering previously hidden vulnerabilities and strengthening the resilience of the global financial architecture.

Across all countries and configurations, GPT-5-mini consistently demonstrates lower average macroeconomic shock severity and linear CVaR multiples compared to Llama-3.1-8B-Instruct, indicating superior risk management.
Across all countries and configurations, GPT-5-mini consistently demonstrates lower average macroeconomic shock severity and linear CVaR multiples compared to Llama-3.1-8B-Instruct, indicating superior risk management.

A Paradigm Shift: Scenario Generation via Large Language Models

Current macro-financial scenario generation relies heavily on methods such as historical simulation, vector autoregression, and agent-based modeling; however, these techniques often lack the flexibility to efficiently explore a wide range of plausible future states and can be computationally expensive to recalibrate or extend. This work introduces an alternative approach leveraging Large Language Models (LLMs) to generate scenarios, offering increased scalability due to the LLM’s inherent ability to extrapolate and synthesize information. The proposed method aims to overcome limitations of traditional techniques by enabling the rapid creation of diverse, yet economically coherent, scenarios based on textual prompts and structured data inputs, allowing for more dynamic and responsive risk assessment and stress testing.

The methodology integrates Large Language Models (LLMs) with structured macroeconomic datasets to enhance the realism and economic validity of generated scenarios. This coupling moves beyond purely textual generation by anchoring LLM outputs to quantifiable economic fundamentals such as GDP, inflation rates, unemployment figures, and interest rates. By conditioning the LLM on this structured data, the system avoids generating scenarios that are economically implausible or inconsistent with established relationships. Specifically, macroeconomic time series data and statistical relationships are incorporated as contextual inputs, influencing the LLM’s predictive capabilities and ensuring that generated scenarios adhere to observed economic behaviors. This data-driven approach is critical for producing robust and reliable macro-financial simulations.

The scenario generation method employs a Retrieval-Augmented Generation (RAG) pipeline to enhance the Large Language Model’s (LLM) output with factual macroeconomic data. This pipeline functions by first identifying data relevant to the LLM’s prompt using FAISS, a library optimized for efficient similarity search within large datasets. FAISS enables rapid retrieval of pertinent economic indicators and historical data. This retrieved data is then incorporated into the prompt provided to the LLM, effectively “grounding” the generated scenarios in real-world economic fundamentals and reducing the likelihood of generating unrealistic or unsupported projections. The RAG architecture ensures that the LLM leverages both its inherent language capabilities and external, structured data to produce more reliable and auditable macro-financial scenarios.

The scenario generation method employs both GPT5mini and Llama318BInstruct to increase the diversity of generated macro-financial simulations. Utilizing two distinct LLMs allows for exploration of a wider range of potential economic pathways compared to a single model approach. Quantitative analysis demonstrates the stability of the resulting scenarios, with linear Value at Risk (VaR) and Conditional Value at Risk (CVaR) multiples consistently ranging between 1.46 and 1.48. This narrow range indicates a consistent level of risk across generated simulations, facilitating auditable and reliable stress-testing and risk assessment.

Enabling news retrieval for Portfolio A via GPT-5-mini with RAG widens the upper tail of the CVaR distribution, corroborating the statistically significant, albeit small, impact of news on portfolio risk as detailed in Table 9.
Enabling news retrieval for Portfolio A via GPT-5-mini with RAG widens the upper tail of the CVaR distribution, corroborating the statistically significant, albeit small, impact of news on portfolio risk as detailed in Table 9.

Translating Macroeconomic Shocks to Quantifiable Financial Risk

A Linear Factor Channel, integrated with Principal Component Analysis (PCA), is utilized to translate macroeconomic shocks – generated by the Large Language Model (LLM) – into quantifiable portfolio risk metrics. The LLM outputs scenarios representing potential economic shifts, which are then processed by PCA to reduce dimensionality and isolate the most significant shocks. This dimensionality reduction identifies the principal components driving risk, allowing for a focused analysis. The Linear Factor Channel then maps these principal components to established financial risk measures, such as Conditional Value at Risk (CVaR), providing a direct link between simulated macroeconomic events and their projected impact on portfolio performance. This process ensures that risk assessments are grounded in the specific economic shocks generated by the LLM, facilitating a more granular and actionable understanding of portfolio vulnerabilities.

Principal Component Analysis (PCA) is utilized to decrease the number of macroeconomic variables required for risk assessment while retaining the most relevant information. By transforming the original set of correlated variables into a smaller set of uncorrelated principal components, PCA identifies the directions of maximum variance in the data, effectively isolating the shocks with the greatest potential impact on portfolio risk. The contribution of each shock is then quantified by examining the loading of each original variable onto the identified principal components, allowing for a focused analysis on the drivers of risk and enabling prioritization of risk mitigation strategies. This dimensionality reduction simplifies the modeling process and improves computational efficiency without sacrificing the accuracy of the risk assessment.

The translation channel systematically connects macroeconomic scenarios, generated via the Large Language Model, to established financial risk metrics. This linkage is achieved by mapping scenario outputs to key risk factors and calculating their impact on portfolio performance. Specifically, the channel facilitates the quantification of portfolio vulnerability by directly translating qualitative scenario descriptions into quantifiable measures such as Conditional Value at Risk (CVaR). This process allows for a comprehensive risk assessment, moving beyond hypothetical scenario analysis to provide concrete data on potential financial losses under various economic conditions, and enables the calculation of linear CVaR multiples ranging from 1.13 to 1.23.

The analysis yields Conditional Value at Risk (CVaR) multiples ranging from 1.13 to 1.23 across tested configurations, providing a quantifiable measure of portfolio risk sensitivity to macroeconomic scenarios. These multiples represent the ratio of CVaR under generated conditions to a baseline CVaR, indicating the magnitude of risk increase. This range suggests a consistent, albeit moderate, amplification of portfolio risk across different macroeconomic simulations. The resulting metrics are directly applicable for risk reporting and portfolio optimization, enabling stakeholders to assess and manage potential losses under stressed conditions. Further granularity within this range is determined by the specific LLM configuration and the macroeconomic variables included in the scenario generation process.

Analysis of Portfolio A reveals that tail risk is driven by the interplay of macroeconomic shocks, regime shifts, and portfolio composition, rather than being solely attributable to inflation, across volatility, linear, and nonlinear translation channels.
Analysis of Portfolio A reveals that tail risk is driven by the interplay of macroeconomic shocks, regime shifts, and portfolio composition, rather than being solely attributable to inflation, across volatility, linear, and nonlinear translation channels.

Ensuring Reproducibility and Validation: A Cornerstone of Reliable Results

A Snapshot Replay strategy is central to ensuring the consistent reliability of reported results. This approach involves meticulously freezing the state of data and analytical models at critical junctures in the research process. By capturing these ‘snapshots’, researchers establish a fixed baseline against which future analyses can be compared, effectively insulating the results from the potentially disruptive effects of evolving datasets or model updates. This practice not only bolsters the reproducibility of findings – allowing independent verification – but also enhances transparency by providing a clear, immutable record of the analytical environment at a specific point in time. The strategy facilitates a robust audit trail, promoting confidence in the validity and dependability of the presented conclusions and enabling consistent performance across different computational environments.

Maintaining consistent results across time is paramount in complex modeling, and a robust system for data and model versioning directly addresses this need. By meticulously freezing data at critical analysis junctures, researchers and practitioners can effectively recreate past results regardless of subsequent updates to the underlying data or model architecture. This capability isn’t merely about technical reproducibility; it fundamentally builds trust in the findings. When analyses can be demonstrably re-executed and validated, stakeholders gain confidence in the reliability and integrity of the conclusions, fostering transparency and enabling more informed decision-making. The ability to consistently generate the same outputs from the same inputs, even as the analytical environment evolves, is therefore a cornerstone of responsible and impactful modeling practices.

Model calibration serves as a crucial refinement process, systematically adjusting the model’s output to achieve congruence with both observed historical data and the informed perspectives of subject matter experts. This isn’t merely about achieving a good fit to past events; it’s about ensuring the model’s predictions are realistically grounded and reflect established knowledge within the field. Through techniques like backtesting and expert elicitation, discrepancies between model forecasts and reality are identified and addressed, leading to a more accurate and reliable predictive capability. The resulting calibrated model demonstrably improves the trustworthiness of its outputs, fostering greater confidence in its application for forecasting and risk assessment, and ultimately delivering more informed decision-making.

Risk decomposition offers a detailed understanding of where potential vulnerabilities originate, moving beyond aggregate risk measures to pinpoint specific contributing factors. This granular approach allows for targeted risk mitigation strategies and improved resource allocation, ultimately strengthening overall risk management. Recent analyses demonstrate the efficacy of this technique, revealing that retrieval processes contribute minimally to Value at Risk (VaR) and Conditional Value at Risk (CVaR), impacting these measures by only 1-2%. Furthermore, the methodology showcases a dispersion of 2.4-3.6 within a simulated macro shock environment, indicating resilience and a controlled response to broad economic fluctuations and providing confidence in the model’s stability under stress.

Linear CVaR multiples for Portfolio A, averaged across all model, RAG, and news configurations in the deterministic run, demonstrate performance relative to a historical-bootstrap baseline (see Table 4).
Linear CVaR multiples for Portfolio A, averaged across all model, RAG, and news configurations in the deterministic run, demonstrate performance relative to a historical-bootstrap baseline (see Table 4).

The pursuit of robust financial risk simulation, as detailed in the paper, demands a foundation built on precisely defined scenarios. Without formal articulation of potential economic shocks, any stress test remains susceptible to ambiguity and unverified assumptions. This aligns with the observation of Blaise Pascal: “Doubt is not a pleasant condition, but certainty is absurd.” The hybrid prompt-RAG pipeline presented acknowledges this inherent uncertainty by generating counterfactuals grounded in both structured data and textual information. The methodology doesn’t claim absolute predictive power, but rather offers a logically consistent, auditable framework for exploring a range of plausible, yet challenging, macroeconomic conditions – a demonstrable attempt to move beyond mere operational testing toward a more mathematically defensible approach to risk assessment.

What Lies Ahead?

The demonstrated pipeline, while functional, merely sketches the boundary of a far more ambitious problem. The generation of plausible macro-financial stress scenarios is not simply a matter of clever prompting or even extensive retrieval-augmentation. The core limitation resides in the inherent stochasticity of large language models, and their susceptibility to producing narratives that appear coherent, yet lack the internal consistency required for rigorous risk simulation. The present work addresses this with RAG, but a truly robust system necessitates a formal verification of generated scenarios against established economic invariants – a task currently beyond the capabilities of the models themselves.

Future research must focus on embedding economic principles directly within the generative process, perhaps through the development of constrained decoding algorithms or the integration of symbolic reasoning engines. The current reliance on post-hoc auditing, while necessary, is fundamentally unsatisfactory; it treats the model as an oracle whose pronouncements must be checked, rather than a tool grounded in demonstrable truth. Asymptotic analysis of the error bounds inherent in LLM-generated scenarios is also critical; simply achieving ‘realistic’ outputs is insufficient when dealing with systemic risk.

Ultimately, the success of this line of inquiry hinges not on scaling model size, but on achieving a deeper understanding of the relationship between linguistic plausibility and economic validity. The pursuit of elegant, provable algorithms remains the only path toward truly reliable stress testing, a principle often lost amidst the current enthusiasm for empirical performance on limited datasets.

Original article: https://arxiv.org/pdf/2512.07867.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-10 18:31