Consumption and Savings with Large Language Model Agents, with Michael R. Douglas
In a canonical consumption-savings model with aggregate productivity and individual employment risk, standard fully rational agents evaluate contingencies and make decisions by following prescriptions from economic theory. This paper replaces them with agents whose intelligence and behavior are powered by Large Language Models, richly parameterized neural networks trained on vast amounts of text. The performance of such LLM agents is in many ways more similar to that of capable but imperfect human beings than their perfectly rational theoretical counterparts. They demonstrate reasonable economic behavior coupled with systematic biases in their decisions and beliefs. The linguistic capabilities of LLMs also permit extensive inter-agent communication. The general approach is additionally validated in a game-theoretic setting.
Ignorance and Indifference: Decision-Making in the Lab and in the Market
(Supplement)
Economic agents live in a dynamic, perpetually changing environment. The ensuing scarcity of information is tackled with invariant ignorance priors, forcing said agents to overweight the chances of less frequent events. Beliefs are updated from diverse information sources, each source assigned an appropriate degree of confidence, and with directly sampled information serving as a canonical metric. Empirically, these mechanisms allow to rationalize and reproduce (i) the prospect theory’s probability transformations and the choices from experienced, described, or ambiguous probabilities (while also accounting for the effect of dynamic belief updating) documented earlier in laboratory experiments, as well as (ii) the equity premium/risk-free rate levels and equity/dividend volatility scaling observed in financial markets.
Machine Learning the Gravity Equation for International Trade, with Michael R. Douglas
Machine learning (ML) is becoming more and more important throughout the mathematical and theoretical sciences. In this work we apply modern ML methods to gravity models of pairwise interactions in international economics. We explain the formulation of graphical neural networks (GNNs), models for graph-structured data that respect the properties of exchangeability and locality. GNNs are a natural and theoretically appealing class of models for international trade, which we demonstrate empirically by fitting them to a large panel of annual-frequency country-level data. We then use a symbolic regression algorithm to turn our fits into interpretable models with performance comparable to state of the art hand-crafted models motivated by economic theory. The resulting symbolic models contain objects resembling market access functions, which were developed in modern structural literature, but in our analysis arise ab initio without being explicitly postulated. Along the way, we also produce several model-consistent and model-agnostic ML-based measures of bilateral trade accessibility.
Modeling Multivariate Time Series in Economics: from Auto-Regressions to Recurrent Neural Networks
The modeling of multivariate time series in an agnostic manner, without assumptions about underlying theoretical structure is traditionally conducted using Vector Auto-Regressions. They are well suited for linear and state-independent evolution. A more general methodology of Multivariate Recurrent Neural Networks allows to capture non-linear and state-dependent dynamics. This paper takes a range of small- to large-scale Long Short-Term Memory MRNNs and pits them against VARs in an application to US data on GDP growth, inflation, commodity prices, Fed Funds rate and bank reserves. Even in a small-sample regime, MRNN significantly outperforms VAR in forecasting out-of-sample. MRNN also fares better in interpretability by means of impulse response functions: for instance, a shock to the Fed Funds rate variable generates system dynamics that are more plausible according to conventional economic theory. Additionally, the paper shows how, due to its inherent non-linearity, MRNN can discover (in an unsupervised manner) different macroeconomic regimes. Utilizing its state dependence, MRNN may also be a useful tool for policy simulations under practically relevant economic conditions (such as Zero Lower Bound).
Thinking on Their Feet: Along Main Street
This paper considers the problem of learning and decision-making in a dynamic stochastic economic environment by agents subject to information processing constraints. An agent endogenously chooses to operate in terms of a simplified model of the economy, which implies: a delayed, if at all, updating of the estimates of evolving states/random variables' conditioning parameters; as well as the entropy reduction, or even its complete ``folding'' that drops the less important variables from the agent's approximating model. Specifically, parameter learning is implemented relying on computational complexity theory, which produces a constrained version of the standard Kalman filter. The latter leads to a less than one-for-one reaction to the newly observed information, without the need to postulate e.g. habit formation; which is responsible for an underreaction to permanent parameter changes (``stickiness''), as well as for an overreaction to transitory shocks (``overshooting''). In a standard stochastic growth model with government transfers, such agents may fail to realize that a fiscal expansion now necessitates a compensatory fiscal contraction later, which implies the effectiveness, in certain sense, of the fiscal stimulus policy (albeit at the expense of efficiency losses) and a violation of the Ricardian equivalence. Numerical simulations suggest high fiscal multipliers, with the effects relatively stronger at times of economic recession. Being the outcomes of endogenous choices of rational agents, these results are immune to the Lucas critique.
Log, Stock and Two Simple Lotteries
(Supplement I, Supplement II)
This paper studies the problem of decision-making under risk by agents whose information-processing abilities may be limited. The theoretical approach taken here relies on economic laboratory experiments, neuroscientific findings, and information-theoretic formalism. The derived mechanism for processing information can be built into the Lucas tree model, a general equilibrium macro-finance workhorse. It simplifies, though also distorts the agent's subjective perspective of the objective stochastic environment. The former converges to the latter when information-processing capacity is sufficiently large (inducing standard ``rational expectations''); but in a more realistic case of bounded capacity, certain rational biases of perception emerge endogenously. The most non-trivial one is categorization: it implies dropping from consideration the less important principal dimensions (``dispersion folding'') and amplifying the random variables' interdependencies (``correlation inflation''). This result helps explain the existence of and variations in the practice of style investing in financial markets.
Thinking on Their Feet: Along Wall Street
This paper studies decision-making under computational complexity constraints within a dynamic non-i.i.d. stochastic environment. An application to financial markets generates underreaction (``inertia''), as well as overreaction (``momentum'') and ensuing ``excess'' volatility.
Solvency Strains and the Long-End of the Yield Curve.
With Francesco Garzarelli and Aqib Aslam. 2010. Goldman Sachs, Fixed Income Monthly, September: 5--7.
Finding `Fair Value' for the Ukrainian Currency.
With Malachy Meechan. 2009. Goldman Sachs, Global Viewpoint, 09/10.
Larger Supply No Threat to Bonds.
With Francesco Garzarelli. 2008. Goldman Sachs, Global Economics Weekly, 08/38: 2--5.
Larger Supply No Threat to Bonds.
2008. Goldman Sachs, Fixed Income Monthly, October: 5--7.
The Family of GS Curves.
2008. Goldman Sachs, Global Viewpoint, 08/16.
Linking the Term Structures of Interest Rates and Macroeconomic Expectations --- GS Yield Curve Valuation Model.
2007. In Jim O'Neill, Jens Nordvig, and Thomas Stolper, eds., The Foreign Exchange Market: 2007. Goldman Sachs: 45--55.
Sudoku Gets Bigger and Forward-Looking!
With Francesco Garzarelli and Michael Vaknin. 2007. Goldman Sachs, Global Viewpoint, 07/24.
Global Bonds: No Obvious Safe Haven.
With Francesco Garzarelli. 2007. Goldman Sachs, Global Viewpoint, 07/07.
Partisan Differences in Economic Outcomes and Corresponding Voting Behavior: Evidence from the U.S.
Public Choice, 120(1-2): 169--189.