Essays in Applied Microeconomics

Abstract

The benchmark model of economics, expected utility theory, assumes that individuals manage uncertainty by forming beliefs about all relevant events and that these beliefs satisfy the probability axioms. There is ample scope for individuals to come to radically different conclusions on the likelihood of events. The first chapter of this thesis examines differences between individuals in how their beliefs evolve in the context of stock returns. In this chapter, which is joint with Hans-Martin von Gaudecker, we analyse a long panel of households' stock market beliefs to gain insights into the nature of the levels, dynamics, and informativeness of these expectations. In a first step, we classify respondents into one of five groups based on their beliefs data alone. In a second step, we estimate models of expectations at the group level so that belief levels, volatility, and response to information can vary freely across groups. At opposite extremes in terms of optimism, we identify pessimists who expect substantially negative returns and financially sophisticated individuals whose expectations are close to the historical average. Two groups expect average returns around zero and differ only in how they respond to information: Extrapolators who become more optimistic following positive information and mean-reverters for whom the opposite is the case. The final group is characterised by its members being unable or unwilling to quantify their beliefs about future returns. <br /> Expected utility theory cannot fully account for how individuals choose under uncertainty. In some settings, observed decisions of a sizeable fraction of individuals cannot be explained with any probabilistic belief. This suggests that heterogeneity in how individuals manage uncertainty goes beyond beliefs, extending to what has become known as ambiguity attitudes. Analysis of these attitudes for natural events is the subject of the second chapter, which is joint with Hans-Martin von Gaudecker and Christian Zimpelmann. We analyse the stability and distribution of ambiguity attitudes using a representative sample. We employ four waves of data from a survey instrument with high-powered incentives. Structural estimation of random utility models yields three individual-level parameters: Ambiguity aversion, likelihood insensitivity or perceived level of ambiguity, and the variance of decision errors. We demonstrate that these parameters are very heterogeneous but fairly stable over time and across domains. These contexts span financial markets, our main application, and climate change. The ambiguity parameters are interdependent in their interpretation and the precision of their estimates depends on decision errors. To describe heterogeneity in these three dimensions, we adopt a discrete classification approach. A third of our sample comes rather close to the behaviour of expected utility maximisers. Half of the sample is characterized by a high likelihood insensitivity, with thirty per cent ambiguity-averse and twenty per cent making ambiguity-seeking choices for most events. For the remaining eighteen per cent, we estimate sizeable error parameters, which implies that no robust conclusions about their ambiguity attitudes are possible. Predicting group membership with a large number of observed characteristics shows reasonable patterns. <br /> The difficulty of finding reliable probabilities for uncertain events raises the question how individuals aspiring to be rational should approach it. One intriguing proposal is to turn to prediction markets or bookmakers. In the third chapter, I examine betting odds, which are often seen as a credible source of predictions for future events in sports, politics, and entertainment. Who will win Wimbledon, who will be the next US President and which movie will win Best Picture at the Oscars? Betting odds, offered by bookmakers or by traders in prediction markets, typically exist for answers to each of these questions and can be turned into implied probabilities. Are these well-calibrated in the sense that they indicate the empirical frequency of outcomes? Do they incorporate publicly available information that might be relevant for prediction? Using a large sample of ATP tennis matches, I investigate these questions with machine learning methods that combine model-based ratings of player strength with a large number of other player features to estimate probabilities. I find that, in almost all settings, implied probabilities are very well calibrated and can be regarded as probabilities of events that condition on an information set containing publicly available statistics on players and matches. They reduce the error of the best prediction model by around 1.3% in terms of negative log-likelihood.