Deep Dive

Deep Dive

Apr 30, 2025

Apr 30, 2025

AI in Inflation Forecasting: How Large Language Models Are Changing Economic Predictions

AI in Inflation Forecasting: How Large Language Models Are Changing Economic Predictions

Watch Video

Watch Video

Watch Video

Predicting inflation is one of the most persistent and challenging economic puzzles. It affects everyone—from families meticulously planning their household budgets to central banks making critical decisions on national interest rates. The variables influencing inflation are vast, changes can be sudden, and the outcomes have a profound impact on daily life and the broader economy. Now, groundbreaking new research from the Federal Reserve Bank of St. Louis suggests that artificial intelligence—specifically, sophisticated large language models (LLMs)—may provide a valuable new lens for understanding and anticipating this elusive target.

The Limitations of Traditional Inflation Forecasting Methods

Historically, the task of inflation forecasting has relied heavily on a combination of surveys of expert opinion, individual judgments from seasoned economists, and complex mathematical or statistical models. Prominent examples include the Survey of Professional Forecasters (SPF) or the intricate DSGE (Dynamic Stochastic General Equilibrium) models used by many central banks. While these traditional tools have provided valuable insights, they can also be time-consuming to develop and maintain, costly to implement, and often struggle to accurately predict inflation during periods of rapid economic change or unprecedented shocks.

Introducing the AI Forecaster: Large Language Models for Economic Prediction

The innovative research conducted by Federal Reserve economists Miguel Faria e Castro and Fernando Leibovici set out to explore whether a large language model—in this specific case, Google's PaLM model (specifically, the Bison 001 version)—could offer more accurate and timely inflation predictions. LLMs are a type of AI trained on massive volumes of text data, enabling them to learn and detect subtle patterns, absorb contextual information, and understand nuances in language that humans might overlook or that traditional models aren't designed to capture.

Simulating Past Economic Conditions to Predict Future Inflation

The researchers employed a clever methodology. They didn’t simply ask the PaLM model, "What’s next for inflation?" in the present day. Instead, they utilized "conditional forecasts." This involved prompting PaLM to act as though it was operating at a specific point in the past—for instance, February 2020—and then asking it to predict future inflation using only the economic information and data that would have been available at that precise historical moment. This approach allowed them to simulate real-world forecasting conditions rigorously, without giving the LLM an unfair advantage by exposing it to information it wouldn't have had at the time of the simulated forecast.

A Closer Look at the AI Model's Inflation Forecasting Results

The PaLM model provided detailed inflation forecasts and, interestingly, often included accompanying economic reasoning. Its outputs frequently discussed factors such as Federal Reserve policy stances, global oil prices, unemployment rates, and other relevant contextual economic signals. Crucially, when the AI's predictions were tested across various historical time periods and for different forecast horizons (e.g., short-term vs. medium-term), PaLM consistently outperformed the Survey of Professional Forecasters (SPF) in most instances. This outperformance was particularly notable during the volatile economic years of the COVID-19 pandemic.

Differences in Forecast Behavior: AI vs. Human Experts

Both the human experts surveyed in the SPF and the PaLM AI model showed a tendency to anticipate that inflation would eventually revert to the Federal Reserve’s stated 2% target. However, a key difference emerged: PaLM exhibited a slower reversion to this target. This suggests that the AI model might be weighing incoming economic signals differently than human forecasters or could be less anchored by explicit policy targets, thereby offering a potentially more cautious or data-driven perspective on the inflation outlook.

Addressing Robustness and Reliability in AI-Generated Forecasts

The study proactively addressed common criticisms leveled against AI models, such as concerns about randomness in outputs or sensitivity to specific prompts. Researchers rigorously tested various phrasings for their prompts, adjusted the model’s internal randomness settings (often referred to as "temperature"), and generated multiple forecasts for each scenario to establish reliable average predictions. Furthermore, they conducted tests to ascertain whether the model could genuinely distinguish between making a real-time forecast based on limited past data versus simply recalling known historical outcomes, and the results suggested that it could indeed perform conditional forecasting effectively.

Implications for the Future of Economic Forecasting

This pioneering research from the St. Louis Fed hints at a potentially seismic shift in how economic forecasts might be conducted in the future. Large language models offer a method for forecasting that could be significantly faster, cheaper, and more scalable than many traditional approaches, especially for economic variables where comprehensive historical data is limited or costly to obtain. One can imagine a future where businesses of all sizes, local government entities, or even individuals could access sophisticated, real-time economic forecasts generated from current media reports, policy announcements, and widespread economic commentary.

Final Thought: AI as a Complement to Human Economic Expertise

While artificial intelligence is not on the verge of replacing human economists entirely, this study compellingly demonstrates that it can significantly complement and enhance traditional forecasting methods. As LLMs continue to improve in their capabilities and become more widely accessible, they are likely to play an increasingly essential role in how we understand, anticipate, and respond to economic changes—possibly even democratizing access to sophisticated economic insights that were once the domain of a select few.

Frequently Asked Questions (FAQs) about AI and Inflation Forecasting

  1. What is the main takeaway from the St. Louis Fed's research on AI and inflation?
    The study found that Google's PaLM large language model was often more accurate than traditional human forecasters (as represented by the Survey of Professional Forecasters) at predicting inflation, particularly for medium-term forecast horizons and during volatile economic periods.

  2. What are conditional forecasts in the context of this AI study?
    Conditional forecasts are predictions made by the AI model where it is instructed to act as if it is at a specific point in the past. The model is then restricted to using only the information and data that would have been available up to that simulated historical date.

  3. How did the researchers test the AI model for reliability and consistency?
    Researchers tested the AI's reliability by varying the prompts given to the model, controlling for its internal randomness settings, and generating and averaging multiple forecast outputs for each scenario to ensure the results were consistent and not due to chance.

  4. Does this research mean AI will replace traditional economists?
    Not yet. The study suggests that AI, particularly LLMs, shows strong potential as a powerful tool to augment and improve the forecasting methods and models currently used by economic experts, rather than replacing them outright.

  5. Could this type of AI be used for forecasting other economic indicators besides inflation?
    Yes, it is plausible. LLMs could be particularly useful for forecasting in areas where traditional data is scarce, costly to collect, or where qualitative information and sentiment play a significant role, potentially extending to other economic indicators.

Hashtags:
#InflationForecasting #AIinEconomics #FederalReserve #StLouisFed #ArtificialIntelligence #GooglePaLM #EconomicForecasts #LLMResearch #FutureOfForecasting #FintechInnovation #MachineLearning #EconomicData #MonetaryPolicy

Subscribe to our Newsletter

Ready to unlock the power of AI for your organization?

Let's discuss how we can partner to achieve your vision.

Address:

Urb. Four Seasons, Los Flamingos Golf,

29679 Benahavís (Málaga), Spain

Contact:

NIF:

ESB44635621

© 2024 Los Flamingos Research & Advisory. All rights reserved

Ready to unlock the power of AI for your organization?

Let's discuss how we can partner to achieve your vision.

Address:

Urb. Four Seasons, Los Flamingos Golf,

29679 Benahavís (Málaga), Spain

Contact:

NIF:

ESB44635621

© 2024 Los Flamingos Research & Advisory. All rights reserved

Ready to unlock the power of AI for your organization?

Let's discuss how we can partner to achieve your vision.

Address:

Urb. Four Seasons, Los Flamingos Golf,

29679 Benahavís (Málaga), Spain

Contact:

NIF:

ESB44635621

© 2024 Los Flamingos Research & Advisory. All rights reserved