As Large Language Models (LLMs) become increasingly integrated into financial systems, understanding their behavioural properties is crucial. Do LLMs conform to the rational expectations paradigm, do they exhibit human-like "animal spirits", or do they instead manifest distinct "machine spirits"? We investigate these questions with a simulated financial market, exploring the behaviour of 15 LLMs spanning a range of sizes, capabilities, and providers. Our results show that LLMs exhibit a spectrum of economic behaviours, from stable coordination on the fundamental value to human-like speculative bubbles. These behaviours are generally inconsistent with the rational expectations hypothesis. We also consider an ecology of heterogeneous agents, a more realistic setting compared to markets with identical LLM agents. These mixed markets can produce outcomes which vary substantially across repeated simulations. Even the most advanced models fail to consistently stabilise the market, with price bubbles sometimes forming despite only a minority of agents naturally forming bubbles. Instead, advanced models in mixed markets adapt their forecasting strategies to the behaviour of other agents. This adaptation can allow them to successfully exploit less sophisticated counterparts and achieve higher profits, but can also contribute to increased market volatility. These findings suggest that the introduction of AI agents into financial markets fundamentally reshapes their ecology. In particular, heterogeneous populations of LLMs can generate endogenous instability, while individual-level adaptation may amplify, rather than mitigate, market volatility.

This study investigates the potential of agent-based modelling to forecast economic crises, addressing the failure of standard macroeconomic models to predict the 2008 financial crisis and capture crisis dynamics. While dynamic stochastic general equilibrium models have incorporated financial frictions, solving them typically requires linearisation around steady states, which suppresses the non-linear feedback loops through which crises emerge. Agent-based models avoid this limitation by numerically simulating heterogeneous agents, preserving non-linear dynamics without approximation. We develop such an agent-based model for the euro area and show that out-of-sample forecasts outperform benchmarks. We further demonstrate that the model can forecast economic crises without exogenous shocks and accurately reproduce crisis dynamics. The model endogenously predicts the onset of the Great Recession, explains the persistence of the sovereign debt crisis, and reproduces the sharp contraction and swift recovery of the COVID-19 recession. The findings suggest that preserving non-linear feedback loops is essential for crisis prediction.

We explore the potential of Large Language Models (LLMs) to replicate human behavior in economic market experiments. Compared to previous studies, we focus on dynamic feedback between LLM agents: the decisions of each LLM impact the market price at the current step, and so affect the decisions of the other LLMs at the next step. We compare LLM behavior to market dynamics observed in laboratory settings and assess their alignment with human participants' behavior. Our findings indicate that LLMs do not adhere strictly to rational expectations, displaying instead bounded rationality, similarly to human participants. Providing a minimal context window i.e. memory of three previous time steps, combined with a high variability setting capturing response heterogeneity, allows LLMs to replicate broad trends seen in human experiments, such as the distinction between positive and negative feedback markets. However, differences remain at a granular level–LLMs exhibit less heterogeneity in behavior than humans. These results suggest that LLMs hold promise as tools for simulating realistic human behavior in economic contexts, though further research is needed to refine their accuracy and increase behavioral diversity.

We study an heterogenous asset pricing model in which different classes of investors coexist and evolve, switching among strategies over time according to a fitness measure. In the presence of boundedly rational agents, with biased forecasts and trend following rules, we study the effect of two types of speculation: one based on fundamentalist and the other on rational expectations. While the first is only based on knowledge of the asset underlying dynamics, the second takes also into account the behavior of other investors. We bring the model to data by estimating it on the Bitcoin Market with two contributions, relying on methods from Machine Learning. First, we construct the Bitcoin Twitter Sentiment Index (BiTSI) to proxy a time varying bias. Second, we propose a new method based on a Neural Network, for the estimation of the resulting heterogeneous agent model with rational speculators. We show that the switching finds support in the data and that while fundamentalist speculation amplifies volatility, rational speculation has a stabilizing effect on the market.

We develop the Canadian behavioral Agent-Based Model (CANVAS) that complements traditional macroeconomic models for forecasting and monetary policy analysis. CANVAS represents a next-generation modeling effort featuring enhancements in three dimensions: introducing household and firm heterogeneity, departing from rational expectations, and modeling price and quantity setting heuristics within a production network. The expanded modeling capacity is achieved by harnessing large-scale Canadian micro- and macroeconomic datasets and incorporating adaptive learning and simple heuristics. The out-of-sample forecasting performance of CANVAS is found to be competitive with a benchmark vector auto-regressive (VAR) model and a DSGE model. When applied to analyze the COVID-19 pandemic episode, our model helps explain both the macroeconomic movement and the interplay between expectation formation and cost-push shocks. CANVAS is one of the first macroeconomic agent-based models applied by a central bank to support projection and alternative scenarios, marking an advancement in the toolkit of central banks and enriching monetary policy analysis.

How do rational and boundedly rational agents interact in a competitive asset market? To answer this question, we build a highly nonlinear asset pricing model where agents hold heterogeneous beliefs. Our model features fully rational forward looking agents versus boundedly rational backward looking agents whose market shares evolve endogenously. This gives rise to chaotic model dynamics which are characterized by complex bubble and crash dynamics, even without any exogenous fluctuations. We show that computational methods can be applied to numerically analyze models combining agents forming rational expectations and agents forming extrapolative expectations, with the possibility of transition between one type of behavior and the other. Not only do we find that boundedly rational agents remain in the market, but document that their effect on price dynamics is even amplified by the behavior of fully rational agents. In their interaction, trend-extrapolators amplify small deviations from fundamentals, while rational agents eventually anticipate market crashes after large bubbles and drive prices back to the fundamental.

This paper brings novel insights into group coordination and price dynamics in complex environments. We implement an overlapping-generation model in the lab, where the output dynamics is given by the well-known chaotic quadratic map. This model structure allows us to study previously unexplored parameter regions where the perfect-foresight dynamics exhibits chaotic dynamics. This paper highlights three key findings. First, the price converges to the simplest equilibria, namely the monetary steady state or the two-cycle, in all markets. Second, we document a novel and intriguing finding: we observe a non-monotonicity of the behavior when complexity increases. Convergence to the two-cycle occurs for the intermediate parameter range, while both the extreme scenarios of a simple stable two-cycle and highly non-linear dynamics (with chaos) lead to coordination on the steady state in the lab. All indicators of coordination and convergence significantly exhibit this non-monotonic relationship in the learning-to-forecast experiments and this non-monotonicity persists in the learning-to-optimize design. Third, convergence in the learning-to-optimize experiment is more challenging to achieve: coordination on the two-cycle is never observed, although the two-cycle Pareto-dominates the steady state in our design.

We study the effect of a "leaning against the wind" monetary policy on asset price bubbles in a learning-to-forecast experiment, where prices are driven by the expectations of market participants. We find that a strong interest rate response is successful in preventing or deflating large price bubbles, while a weak response is not. Giving information about the interest rate changes and communicating the goal of the policy increases coordination of expectations and has a stabilizing effect. When the steady-state fundamental price is unknown and the interest rate rule is based on a proxy instead, the policy is less effective.

We conduct an information-provision experiment within a large-scale household survey on public finance in France, The Netherlands and Italy. We elicit prior opinions via open-ended questions and introduce a measure of macroeconomic policy literacy. A central bank (CB) educational blogpost explaining the mechanics of CB money preceded by a short video clip on public finance can persistently induce less support for monetary-financed proposals and more for fiscal discipline and CB independence, no matter the respondents’ level of policy literacy. However, prior beliefs matter and contradictory information may be polarizing. Additional analysis of our data shows that information affects the respondents’ views by shifting their inflation and tax expectations associated to these policies.

We provide experimental evidence on coordination within large groups that could proxy the atomistic nature of real-world markets. We use a bank run game where the two pure-strategy equilibria can be ranked by payoff and risk dominance and a sequence of public announcements introduces stochastic sunspot equilibria. We find systematic group size effects that theory fails to predict. When the payoff-dominant strategy is risky enough, the behavior of small groups is uninformative of the behavior in large groups: unlike smaller groups of size ten, larger groups exclusively coordinate on the Pareto-inferior strategy and never coordinate on sunspots. (JEL C92, D83, D91, G21)

This paper analyses the post-pandemic inflation dynamics in Canada using a behavioral macroeconomic model of the Bank of Canada. Two crucial behavioral assumptions in line with empirical evidence characterize the model: firms make price-quantity decisions according to a simple heuristic rule and form expectations using adaptive learning. The paper shows that post-pandemic inflation can be explained by these simple heuristics and traced back to the lifting of economic restrictions in mid-2020, triggering demand-pull inflation. This result suggests that the cost-push inflation in 2021 and beyond actually originated from a demand-pull mechanism, making the latter the primary driver of post-pandemic inflation.

Expectations of future returns are pivotal for investors’ trading decisions, and are therefore an important determinant of the evolution of actual returns. Evidence from individual choice experiments with exogenously given time series of returns suggests that subjects’ return forecasts are substantially affected by how they are elicited and by the format in which subjects receive information about past asset performance. In order to understand the impact of these effects found at the individual level on market dynamics, we consider a learning to forecast experiment where prices and returns are endogenously determined and depend directly upon subjects’ forecasts. We vary both the variable (prices or returns) subjects observe and the variable (prices or returns) they have to forecast, with the same underlying data generating process for each treatment. Although there is no significant effect of the presentation format of past information, we do find that markets are significantly more unstable when subjects have to forecast returns instead of prices. Our results therefore show that the elicitation format may exacerbate, or even create, bubbles and crashes in financial markets.

We develop an agent-based model for the euro area that fulfils widely recommended requirements for next-generation macroeconomic models by i) incorporating financial frictions, ii) relaxing the requirement of rational expectations, and iii) including heterogeneous agents. Using macroeconomic and sectoral data, the model includes all sectors (financial, non-financial, household, and a general government) and connects financial flows and balance sheets with stock-flow consistency. The model, moreover, incorporates many features considered essential for future policy models, such as a financial accelerator with debt-financed investment and a complete GDP identity, and allows for non-linear responses. We first show that the agent-based model outperforms dynamic stochastic general equilibrium and vector autoregression models in out-of-sample forecasting. We then demonstrate that the model can help make sense of extreme macroeconomic movements and apply the model to the three recent major economic crises of the euro area: the Financial crisis of 2007-2008 and the subsequent Great Recession, the European sovereign debt crisis, and the COVID-19 recession. We show that the model, due to non-linear responses, is capable of predicting a severe crisis arising endogenously around the most intense phase of the Great Recession in the euro area without any exogenous shocks. By analysing the COVID-19 recession, we further demonstrate the model for scenario analysis with exogenous shocks. Here we show that the model reproduces the observed deep recession followed by a swift recovery and also captures the persistent rise in inflation following the COVID-19 recession.

We introduce Behavioral Learning Equilibria (BLE) into a multivariate linear framework and apply it to New Keynesian DSGE models. In a BLE, boundedly rational agents use simple, but optimal AR(1) forecasting rules whose parameters are consistent with the observed sample mean and autocorrelation of past data. We study the BLE concept in a standard 3‐equation New Keynesian model and develop an estimation methodology for the canonical Smets and Wouters (2007) model. A horse race between Rational Expectations (REE), BLE, and constant gain learning models shows that the BLE model outperforms the REE benchmark and is competitive with constant gain learning models in terms of in‐sample and out‐of‐sample fitness. Sample‐autocorrelation learning of optimal AR(1) beliefs provides the best fit when short‐term survey data on inflation expectations are taken into account in the estimation. As a policy application, we show that optimal Taylor rules under AR(1) expectations inherit history dependence and require a lower degrees of interest rate smoothing than REE.

We develop the first agent-based model (ABM) that can compete with benchmark VAR and DSGE models in out-of-sample forecasting of macro variables. Our ABM for a small open economy uses micro and macro data from national accounts, sector accounts, input–output tables, government statistics, and census and business demography data. The model incorporates all economic activities as classified by the European System of Accounts (ESA 2010) and includes all economic sectors populated with millions of heterogeneous agents. In addition to being a competitive model framework for forecasts of aggregate variables, the detailed structure of the ABM allows for a breakdown into sector-level forecasts. Using this detailed structure, we demonstrate the ABM by forecasting the medium-run macroeconomic effects of lockdown measures taken in Austria to combat the COVID-19 pandemic. Potential applications of the model include stress-testing and predicting the effects of monetary or fiscal macroeconomic policies.

The impact of finite forecasting horizons on price dynamics is examined in a standard infinite-horizon asset-pricing model. Our theoretical results link forecasting horizon inversely to expectational feedback , and predict a positive relationship between expectational feedback and various measures of asset-price volatility. We design a laboratory experiment to test these predictions. Consistent with our theory, short-horizon markets are prone to substantial and prolonged deviations from rational expectations, whereas markets with even a modest share of long-horizon forecasters exhibit convergence. Longer-horizon forecasts display more heterogeneity but also prevent coordination on incorrect anchors - a pattern that leads to mispricing in short-horizon markets.(c) 2022 Elsevier B.V. All rights reserved.

This paper develops a unified analysis of the impacts of production delays on aggregate price fluctuations in a continuous-time cobweb-type model. We find that the time inconsistency between demand and supply due to production delays inherently generates an overshooting effect in prices. Large production delays give rise to price fluctuations, even with a low intensity of choice. In our model, investor beliefs (prices) are updated slowly (quickly), leading prices to build up slowly but crash suddenly. We also provide results consistent with the rational route to randomness of Brock and Hommes (1997) in a continuous-time infinite-dimensional economy.