In an era of increasing economic volatility, accurate risk measurement has become the cornerstone of financial resilience. Economic Scenario Generators (ESGs) – sophisticated models that simulate potential future economic conditions – stand at the core of risk management in financial organizations. Scenarios generated with an ESG allow financial professionals to assess the risk under various economic conditions, spanning from a booming economy to a deep recession. As financial landscapes grow increasingly complex, the quality and reliability of an organization’s Economic Scenario Generator directly impact its ability to navigate uncertainty.
Historically, the world of finance and economics has been at the forefront of utilizing sophisticated mathematical and statistical models coupled with computational power to manage risks. However, despite the rapid advancement of machine learning and generative AI, the financial industry has largely remained entrenched in its traditional methodological frameworks, failing to capitalize on the transformative potential of these emerging technologies.
Economic Scenario Generators, being an essential component of financial risk management, can significantly benefit from cutting-edge methodologies developed in machine learning and artificial intelligence. In this blog, we compare and contrast traditional statistical ESG frameworks with AI-inspired Economic Scenario Generators and discuss how AI-based models can dramatically improve the quality and adaptability of the generated economic scenarios.
Traditional Statistical Economic Scenario Generators
Traditional Economic Scenario Generators (ESGs) rely on a variety of statistical models to capture the inherent uncertainties and complexities of economic systems. The Society of Actuaries published a practical guide that provides a detailed overview of ESGs and their applications. It also describes commonly used statistical methods in traditional ESGs.
Vector Autoregressive Models
The majority of traditional statistical ESGs utilize Vector Autoregressive (VAR) Models.
A standard Vector Autoregressive model analyzes the interdependencies among multiple time series by regressing each variable on its past values and the past values of all other variables in the system. This captures the dynamic relationships and potential feedback effects between them.
Cascading VAR Models
The cascading version of Vector Autoregression models introduces a sequential or hierarchical structure to the modeling process. Instead of estimating all the equations of the VAR system simultaneously, a cascading VAR model estimates the equations in a specific order, where the results from the earlier stages influence the subsequent stages. This framework prioritizes the modeling of leading economic variables before addressing lagging ones, ensuring that earlier-stage dynamics shape the subsequent stages.
Advantages of Traditional Statistical ESGs
Traditional statistical ESG approaches have long been valued for their structured methodologies and ability to provide interpretable economic projections. These established frameworks offer several competitive advantages that have cemented their position in financial risk management:
- Allow incorporation of domain knowledge, such as
- Setting the lag structure of the time series
- Calculation order of the economic variables
- Imposing restrictions on model parameters.
- Ease of interpretation of coefficients, such as the direction and the magnitude of change in one variable given the change in another variable.
- Ease of estimation, especially with a small sample size.
- Well-documented theoretical foundation and regulatory acceptance.
Disadvantages of Traditional Statistical ESGs
While traditional statistical Economic Scenario Generators have served as the backbone of financial risk modeling for decades, their inherent structural limitations increasingly constrain their effectiveness in today’s complex economic landscape. These methodological rigidities create significant blind spots that can compromise scenario quality and comprehensiveness when modeling modern economic dynamics:
- Inability to capture non-linear and complex effects, including level dependency and multi-variable interactions.
- The lag structure is fixed and has to be provided explicitly.
- Regime switches have to be modeled explicitly.
- Requirement of stationarity of the time series. As a result, non-stationary time series have to be transformed to achieve stationarity before feeding them into the model.
AI-based Economic Scenario Generator
The development of advanced deep learning architectures and algorithms, boosted by the increasingly available computational power, led to the creation of generative AI models. While Large Language Models (or LLMs) are the most popular generative models, they are not well-suited for economic scenario generation.
Fortunately, other types of AI model architectures can be utilized to generate economic scenarios. Some popular generative models are Variational Autoencoders (VAEs), Denoising Diffusion Probabilistic Models (DDPMs), and Normalizing Flows. The main idea behind these models is to map a point in a multidimensional normal distribution into an economic scenario. Let’s explore these models in a bit more detail and discuss their unique approaches to generating economic scenarios.
Variational Autoencoder as ESG
VAEs consist of an encoder and a decoder network. During model training, the encoder learns to map the input economic scenarios to a latent probability distribution, often assumed to be a normal distribution. The decoder, in its turn, learns to map these latent points back to the original scenarios.
During generation, a point is sampled from the latent normal distribution, and this sample is then fed into the decoder. The decoder maps this latent point back to the scenario space, generating a new economic scenario. VAEs learn a smooth and continuous latent space, which enables meaningful interpolation between generated scenarios.
Denoising Diffusion Probabilistic Models as ESG
DDPMs work by defining a forward (or diffusion) process that gradually adds noise to economic scenarios until they become pure noise (approximating a normal distribution). The model then learns to reverse this process by iterative denoising to obtain the original economic scenarios. The reverse process is learned by predicting the noise added at each step of the forward process.
To generate a scenario, DDPM iteratively denoises a random point sampled from a multidimensional normal distribution to simulate a new economic scenario. With proper training, DDPMs are capable of generating high-quality economic scenarios.
Normalizing Flows as ESG
Normalizing Flows transform a simple initial probability distribution (like a standard normal distribution) into the complex distribution of economic scenarios through a sequence of invertible mappings. To generate a scenario, a point is drawn from the base normal distribution and then passed through the sequence of invertible transformations. By composing multiple simple invertible functions, Normalizing Flows can model very complex distributions of economic scenarios.
Normalizing Flows can be a computationally effective alternative to DDPMs.
Advantages of AI-based Economic Scenario Generators
The evolution towards AI-based ESGs marks a significant leap forward, offering distinct benefits over the established statistical methodologies. The main advantages of AI-based Economic Scenario Generators are
- Ability to automatically capture linear, lagging, and complex non-linear effects, such as level dependency and interactions between multiple variables.
- Capability to capture regime switches.
- Better simulation of extreme economic scenarios (tail-risk).
- The capacity to work well with high-dimensional data.
- Allows for the incorporation of additional data, including unstructured data such as text.
- Requires only normalization of the time series, which is easy to achieve.
Disadvantages of AI-based Economic Scenario Generators
While AI-based ESGs offer many advantages over traditional ESGs, they introduce a new set of methodological hurdles that must be systematically addressed to achieve a robust Economic Scenario Generator.
- AI-based ESGs may suffer from mode collapse when the model generates only certain types of scenarios, such as only recessionary scenarios.
- May produce fat-tail distributions where extreme scenarios are generated more often than one would expect.
- Requires a large volume of data for reliable model estimation. Historical economic data is usually short, resulting in significant challenges in model fitting.
- Incorporating domain knowledge is not as easy as in traditional ESG models.
- The black-box nature of AI-based ESGs does not allow straightforward interpretation of the model output.
Real-World vs. Risk-Neutral Economic Scenarios
In economic and financial analysis, there are two fundamental frameworks for evaluating economic conditions and pricing assets. The real-world economic scenarios reflect actual probabilities of economic events as they occur in reality. They are used for capital budgeting, stress testing, and forecasting actual outcomes.
In contrast, the risk-neutral economic scenarios are based on a theoretical construct where investors are indifferent to risk. They are primarily used for derivative pricing.
Traditional ESGs can generate both real-world and risk-neutral scenarios. The VAR (or cascading VAR) method mentioned earlier is used for generating real-world economic scenarios. For risk-neutral scenarios, Heath–Jarrow–Morton (HJM) is a commonly used framework to generate the evolution of interest rate curves.
AI-based Economic Scenario Generators can produce high-quality real-world scenarios. To achieve it, they require proper model architecture and special techniques to overcome challenges during the model training and generation.
Can AI-based ESGs generate risk-neutral economic scenarios? While there has been some limited research to generate risk-neutral scenarios with Variational Autoencoders, we haven’t found any such model-agnostic AI methodology.
To generate risk-neutral economic scenarios with AI-based ESGs, we utilize a hybrid approach. It combines traditional risk-neutral frameworks with AI-based Economic Scenario Generators. The hybrid approach ensures that the generated scenarios are risk-neutral and, at the same time, they benefit from the richness and sophistication that AI models can offer.
What AI-generated Economic Scenarios Look Like?
As a picture is worth a thousand words, we illustrate a randomly picked AI-generated economic scenario. Similar to popular AI models generating pictures of realistic-looking gods and landscapes, AI-generated scenarios capture the complex patterns and relationships between macroeconomic variables.
The scenario is generated based on the information available up to December 2024. While this scenario is not a forecast of the economy, it represents one of a thousand possible developments that could occur.
Initially, the economy is doing well with healthy employment and a falling inflation rate supported by decreasing oil prices (WTI). Both real GDP and real disposable income grow at a healthy pace.
Around the halfway of the scenario, the unemployment rate sharply rises, prompting the Fed to cut the interest rate to support the economy. However, the lower rates cause an increase in inflation, which is further amplified by the increase in oil prices. GDP growth slows down, then sharply accelerates, likely driven by government stimulus to support the economy (the stimulus can drive inflation too).
The steady growth of house prices reverses at the end of the scenario, reflecting the volatile economy and unstable disposable income.
Interestingly, the stock market frontruns both the increase and the decrease of unemployment, as it often does. VIX spikes just before the sharp increase in the unemployment rate.
Summary
The increasing uncertainty and complexity of economic reality require adaptable Economic Scenario Generators to assess and navigate the macroeconomic risks. The unsophisticated nature of the traditional ESGs is increasingly limiting the risk managers’ ability to measure the macroeconomic exposure of their portfolio.
AI-based Economic Scenario Generators capture the intricate landscape of macroeconomic patterns and interactions, providing a comprehensive view of potential future economic trajectories.
At Scenarios by AI, we provide AI-generated, coherent economic scenarios covering a wide spectrum of macroeconomic conditions and their inherent complexity. Reach out to discuss how AI-based economic scenarios can help you better measure and manage the financial risks.