Amsterdam node

Leader: Cars H. Hommes

In this node ofUniversityofAmsterdam, CeNDEF, the main subject of entry into the network’s program is the experimental validation of Coordination of heterogeneous expectations. In particular, we will focus on using laboratory experiments with human subjects to study expectation coordination and to validate different expectations hypotheses and learning models.

Individual expectations about future economic variables, such as asset prices or inflation, play an important role in real life as well as in economic theory. Expectations shape the decisions of individuals and, hence, affect the realization of aggregate variables. In turn, these realizations may affect individual expectations. In this way one may think of an economy as an expectation feedback system. A standard approach to modelling agents’ expectations has been to assume rational expectations, when market participants form model-consistent predictions of future outcomes, as if they have full knowledge of the economic law of motion. However, the rational expectation models are not satisfactory, neither in their predictions nor in their unrealistic assumptions of perfect rationality. For example, the strong recent decline of financial markets and the current economic crises are hard to reconcile with the rational model. Moreover, non-rational, heterogeneous expectations are frequently found in laboratory experiments as well as in survey data.

There are many alternatives to assuming rational expectations, but typically the predictions of these models critically depend upon which expectation rules are assumed. Deviating from the rational expectations hypothesis gives many degrees of freedom in modeling expectation formation. In order to discipline this so-called “wilderness of bounded rationality’’ (Sims, 1980) we will study, by running laboratory experiments with paid human subjects, how individuals form expectations, whether or not these individual expectations are coordinated and what is the resulting aggregate behavior in the economy.

Indeed, laboratory experiments with human subjects, where the experimenters have full control over the market environment and economic fundamentals, form an ideal tool to study individual expectations and how their interaction shapes aggregate market outcomes. For example, in experimental asset markets (e.g. Smith et al., 1988) bubbles and crashes in asset prices may arise. It has been suggested that these non-equilibrium phenomena emerge due to the lack of common knowledge of rationality.

While in earlier experimental work expectations often played an indirect role and their formation and effects were not easy to test, more recently, so-called learning-to-forecast experiments (LtFEs) have been designed to study the expectation formation process directly. In these LtFEs (see Hommes, 2011 for an overview), a subject’s only task is to forecast e.g. the price of some asset, for a number, say 50-60, of periods, with the realization of aggregate economic variables in each period determined by (average) individual expectations. LtFEs thus provides clean data on individual expectations as well as aggregate price behavior.

One of the most striking phenomena found in a number of these LtFEs is the coordination of individual expectations: after a short learning phase subjects submit very similar predictions, despite being uninformed about past and current predictions of the others participants, see e.g. Hommes et al. (2005). Moreover, estimation of individual prediction rules suggests (see e.g. Hommes et al., 2005, Heemeijer et al., 2009, and Assenza et al., 2010) that participants often use simple, behavioral rules, where some participants use adaptive expectations, while others extrapolate trends in past prices. Coordination of expectations also depends upon the nature of the expectations feedback structure. Heemeijer et al. (2009), for example, run LTFEs with two perfectly symmetric treatments, one where the feedback between predictions and the realized price is positive, and one where it is negative. In the latter case participants learn to coordinate on the rational expectations equilibrium quickly. In contrast, in the positive feedback treatment, participants also coordinate their expectations but not on the rational expectations equilibrium, leading to persistent deviations from fundamental values. Similar results have been found in Fehr and Tyran (2008), Sutan and Willinger (2009) and Sonnemans and Tuinstra (2010).

Laboratory experiments are very useful for testing different models of expectation formation. Anufriev and Hommes (2009) and Anufriev et al (2010) show that a model of heterogeneous, dynamically evolving expectations, the so-called Heuristic Switching Model (HSM), fits the data from LtFEs nicely. Neither the rational expectations model nor other homogeneous expectations models (e.g., adaptive or trend-following expectations) outperform the HSM. The HSM is successful in explaining coordination of individual expectations across different market settings.

In this node we will run new LtFEs in a number of different finance and macro-economic environments, such as asset pricing models, overlapping generation models and New Keynesian macro-economic settings. We will use the data from these LtFEs to address the following important questions. (i) How do individuals form expectations? (ii) Will coordination of expectations occur, even when there is limited information, or will heterogeneity persist? (iii) When does learning enforce convergence to rational expectations equilibria and when do boundedly rational “learning equilibria” arise? Moreover, the new experiments will provide additional empirical evidence that we will use to improve the fit of the HSM, and to test the model for robustness with respect to the experimental environment. The HSM model will also be tested on empirical time series, e.g., stock prices and inflation. Finally, we will use the experimental data to organize a competition between different theories of expectations and learning and investigate which theory fits the data most accurately.