Session 3: Applied Artificial Intelligence in Macro-Finance

Marginal q—the shadow price of capital—is a sufficient statistic for investment, yet it is unobservable. For fifty years, empirical work has relied on the average Q, a weak predictor of investment. We take a bottom-up approach to seriously put “marginal” back to the q construction. For each firm-year, we use retrieval-augmented generation (RAG) to assemble relevant textual evidence from filings, patents, earnings calls, peer disclosures, and news, then prompt a large language model to identify plausible marginal projects and estimate their market values and costs. The resulting measure, q AI, substantially outperforms average Q and intangible-adjusted total Q in predicting investment. Because we capture the marginal project itself, we show that the investment-q elasticity is larger for projects compatible with existing operations and smaller for acquisition-driven expansion. Decomposing the average Q-marginal q wedge, we find that capitalized rents, omitted intangibles, and financial frictions each contribute, with relative magnitudes varying across industries.

Evaluating optimized AI systems for asset pricing is fundamentally difficult for two reasons. First, models are trained on all data, implying that any backtest or analysis using historical data suffers from look-ahead bias. In addition, markets are reflexive— as investors adopt AI, prices adjust — which may erode the very patterns the AI system was trained to exploit. We introduce a real-time, out-of-sample benchmark designed to sidestep both problems. The benchmark measures how well AI systems can explain contemporaneous stock returns around earnings announcements using only information available at announcement time, including the text of the announcement itself. Applying this benchmark to a range of agentic AI systems — which extract structured signals from earnings call transcripts and optimize over those signals — we find that the best-optimized systems more than double the explained variation in returns relative to standard benchmarks (R2 increasing from 8% to close to 20%). We also discuss the efficiency gains of AI-based optimization compared to traditional machine learning methods. Beyond predictive performance, these systems generate narrative explanations for price movements that yield testable implications for asset pricing theory. We release an SDK and launch an open competition inviting researchers to improve on our results. Saturating this benchmark would represent fundamental progress in understanding how capital markets process firm-level information.

We develop a framework to forecast inflation using high dimensional microdata in a changing environment with structural breaks. We apply the method to microdata on price changes for the United Kingdom, over 1996-2024. Our contribution is threefold. First, we encode the distribution of price changes into quantiles and moments of the distribution, which combine with a machine learning algorithm to forecast inflation. Second, we develop a scan test for a composite null hypothesis that one forecast never outperforms a second. Rejecting the null implies the first forecast outperforms at an ex ante unknown point in time, of unknown duration, for some unknown forecasting horizon. Applying the scan test sequentially, the macro forecast improves on a univariate benchmark, and the micro forecast adds predictive content beyond the combined macro-univariate benchmark in localized episodes, including the early post-2020 inflation surge. This hypothesis is useful in a non-stationary environment where no single modality is expected to dominate. Third, we combine multiple modalities, micro, macro and univariate forecasts, using the fixed shares algorithm: an adaptive machine learning method with dynamic regret guarantees in adversarially non-stationary settings. The fixed shares algorithm outperforms forecasts that do not use microdata, and places significant weight on microdata after 2020. We discuss the implications of our results for models of inflation dynamics.

We present a new approach to formulating and solving heterogeneous agent models with aggregate risk. We replace the cross-sectional distribution with low-dimensional prices as state variables and let agents learn equilibrium price dynamics directly from simulated paths. To do so, we introduce a structural reinforcement learning (SRL) method which treats prices via simulation while exploiting agents’ structural knowledge of their own individual dynamics. Our SRL method yields a general and highly efficient global solution method for heterogeneous agent models that sidesteps the Master equation and handles models traditional methods struggle with, like those with nontrivial market-clearing co ditions. We illustrate the approach in the Krusell-Smith model, the Huggett model with aggregate shocks, and a HANK model with a forward-looking Phillips curve, all of which we solve globally within minutes.

Emerging techniques in computer science make it possible to "brain scan" large language models (LLMs), identify the plain-English concepts that guide their reasoning, and steer them while holding other factors constant. We show that this approach can map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing, and compute their relative importance without reducing performance. We also show that models can be steered to be more or less risk-averse, optimistic, or pessimistic, which allows researchers to correct or simulate biases. The method is transparent, lightweight, and replicable for empirical research in the social sciences.

We measure and analyze inflation uncertainty in the US. We construct a novel composite indicator of inflation uncertainty (CIU) from two components: a news-based measure derived from textual analysis of newspaper articles using large language models and a market-based measure that draws on prices of options on Exchange Traded Funds and commodities. Unlike survey- or inflation-option-based measures, our index is available in real time and extends back to 1926. CIU reveals that inflation uncertainty spiked during the Great Depression, World War II, the 1970s and 1980s, following the Global Financial Crisis, and in the post-pandemic period. We highlight the driving forces behind these fluctuations in uncertainty and analyze their economic consequences. Heightened inflation uncertainty is associated with higher prices of real assets—such as gold, silver, and housing—but with lower prices of nominal assets, including government bonds, corporate bonds, and equities. More-over, we find that increases in inflation uncertainty are followed by declines in private investment and real economic activity.

Since ChatGPT's release in 2022, demand for artificial intelligence (AI)–related skills in finance has grown rapidly, as generative AI drives significant technological changes in both the financial research field and the broader economy. We show that financial occupations are highly exposed to the productivity effects of generative AI, review the literature on the impact of ChatGPT on firm value, and provide directions for future research investigating the impact of this major technology shock. Generative AI also holds great potential as a tool for finance researchers and practitioners: We review and describe innovations in research methods linked to improvements in AI tools, along with their applications. We offer a practical introduction to available tools and advice for researchers in academia and industry interested in using these tools.

Title: Deep Learning for Solving Economic Models

The ongoing revolution in deep learning is reshaping research across many fields, including economics. Its effects are especially clear in solving dynamic economic models. These models often lack closed-form solutions, so economists have long relied on numerical methods such as value function iteration, perturbation, and projection techniques. Unfortunately, these approaches suffer from the curse of dimensionality, which makes global solutions computationally infeasible as the number of state variables increases. Deep learning offers a different approach: flexible tools that solve dynamic economic models by minimizing residuals in equilibrium conditions, and that can handle high-dimensional problems. This development promises to broaden the scope of quantitative economics. I illustrate the approach using the neoclassical growth model.

We study whether AI methods applied to large-scale portfolio holdings data can improve macroprudential financial regulation. We build a graph-based deep learning model tailored to security-level data on the holdings of financial intermediaries. The architecture incorporates economic priors and learns latent representations of both assets and investors from the network structure of portfolio positions. Applied to the universe of non-bank financial intermediaries, covering nearly $40 trillion in wealth, the model substantially outperforms existing approaches in out-of-sample forecasts of intermediary trading behavior, including in crisis episodes. The model has more than ten times the explanatory power for the cross-sectional variation in asset returns during stress events compared to traditional approaches, and it outperforms existing systemic risk metrics at the institution level. Its learned representations show that the holdings network encodes rich, economically interpretable information about fire- sale vulnerability. The architecture is fully inductive, producing informative estimates even when entire asset classes or investors are withheld from training. We embed our empirical approach into a macroprudential optimal policy framework to formalize why these objects matter for policy and welfare. We show that even in an equilibrium environment subject to the Lucas critique, the predictive information from the model improves welfare by sharpening the cross-sectional targeting of policy interventions, and we demonstrate a complementarity between prediction and structural knowledge.

This paper investigates the impact on financial market stability of AI planning, the core technology enabling agentic AI systems to pursue long-term objectives by anticipating how current actions affect future payoff-relevant states. We develop a dynamic trading framework comprising positive-feedback investors, constrained arbitrageurs, and oligopolistic informed speculators capable of intertemporal coordination. Specifically, these speculators trade aggressively in tandem to generate  (negative) bubbles and subsequently unwind their positions in a coordinated manner to extract profits. This intertemporal coordination diverges from traditional collusion because it faces two fundamental obstacles: time inconsistency, as coordinated plans become incentive-incompatible once a large bubble forms, and weak punishment, as deviations are difficult to penalize in the absence of a significant bubble. We characterize equilibria featuring the coordinated creation of manipulative and exploitative   bubbles. Through simulation experiments, we demonstrate that AI speculators, trained using reinforcement learning algorithms with explicit planning modules, autonomously discover and implement intertemporal collusive strategies based on compounded price-trigger rules, coordinating without communication or shared intent. When feedback trading is strong, these AI speculators dynamically converge on destabilizing strategies that create and exploit bubbles, manipulate feedback traders, and significantly amplify market fragility.