Central Bank Transparency with (semi-)public Information: Laboratory Experiments

Highlights

• Two laboratory experiments are run to test central bank partial transparency strategies: “fragmented information” strategy and “partially hidden information” strategy.

• We provide original evidence that the “fragmented information” strategy outperforms the “partially hidden information” strategy in terms of social welfare, as central banks can better control the mean squared action error and the dispersion of agents’ behavior with the former strategy.

• The central bank can alleviate agents’ overreaction to public information by one or other form of partial transparency.

• The distribution of the weights assigned on the (semi-)public signal emphasizes heterogeneity of behavior entailed by boundedly rational reasoning.

• Learning mechanisms to improve performance and behavior differs with respect to the game played by the subjects.

Abstract

Strategies of public announcements pose challenges to central banks. Theory shows that full transparency is not always good. In this paper, we propose to assess two forms of partial public disclosure by central banks (“fragmented information” and “partially hidden information” strategies) in two beauty contest games, as well as a scenario where public information is fully disclosed. Based on laboratory experiments, we offer original evidence that the “fragmented information” strategy outperforms the “partially hidden information” strategy in terms of social welfare, as central banks can better control the mean squared distance of agents’ actions from the true state of fundamentals (i.e., Mean squared action error) and the dispersion of agents’ behavior with “fragmented information”, while both partial transparency strategies similarly alleviate agents’ overreaction to fully disclosed public information. We also find that divergence from the Nash equilibrium emphasizes heterogeneity of behavior that is entailed by boundedly rational reasoning, especially in early periods of a game. Further, we build on choice reinforcement and belief-based learning models to better understand how subjects learn over time to improve their performance. How well those learning models fit the data depends on the game played by the subjects.

Introduction

Worldwide central banks have availed with many transparent practices on monetary policy and financial stability. Full transparency in central banks’ public announcements may be over-attractive to economic agents, whereas the latter forego the benefits of more (potential) valuable sources of information, such as their (independent) private information. This paper is located under a coordination game with heterogeneous signals, wherein deciding on uncertain economic fundamentals is akin to the agent's own beliefs and to the beliefs about her opponents’ uncertain beliefs.¹ This is the primary feature of a Keynesian beauty contest (Keynes, 1936), exemplified by Morris and Shin (2002) who argue that overweighting public information may deteriorate social welfare.^2,3

The inherent result of Morris and Shin raises numerous “for” and “against” responses.⁴ A first worthwhile branch of theory focuses on the value of a transparent against an opaque regime (e.g., Geraats, 2002; Gosselin et al., 2008; Walsh, 2008). With respect to this branch, central bank transparency boils down to a dichotomous decision of whether to communicate or not public information. The second branch goes beyond the extreme cases and advances the question of how central banks disclose public announcements. Within this line of research, a significant body of theoretical literature challenges the overreaction issue raised in the canonical paper of Morris and Shin (2002) and recommends partial transparency strategies as alternative to full transparency. For example, Cornand and Heinemann (2008) suggest that overreaction to the shared news is allayed if public authorities hide information from a subset of economic agents. We call this strategy “partially hidden information”. With this strategy, we distinguish informed (i.e., public information users) and uninformed (i.e., public information non-users) agents. An equally attractive strategy is attained with a “fragmented information” strategy (see Morris and Shin, 2007; Trabelsi, 2013). Under this strategy, the population is divided into $n$ equal subsets, each subset shares different public information. In the rest of the text, the common signal is public in a full disclosure scenario and semi-public in both the partially hidden and fragmented information scenarios.⁵

Laboratory experiments have been extensively employed to study central bank transparency and communication in various contexts.⁶ We focus our discussion on papers whose experimental design was guided by the theoretical framework of Morris and Shin (2002). For instance, Cornand and Heinemann (2014) and (Shapiro et al., 2014) attribute overreaction to public information to a high value of the coordination motive arising from strategic complementarity.⁷ Now, if a high coordination motive is accompanied with more precise public signal compared to the private one, then agents attach again excessive weight to the former signal (Baeriswyl and Cornand, 2016). The relationship between overreaction to public information and its precision is also discussed in Dale and Morgan (2012). Beforehand, a laboratory asset market study by Middeldorp and Rozenkraz (2011) shows that agents overreact to public announcements whenever the private signal is acquired at a cost.

Testing overreaction in laboratory implies the essential assumption of a full transparency strategy, which is a common feature in the above papers. The novelty of the proposed study is to test partial transparency strategies. Unlike the earlier papers, we neither assume a costly private information like Middeldorp and Rozenkraz (2011), nor consider a variation of signals’ accuracy (see Dale and Morgan, 2012; Baeriswyl and Cornand, 2016) or of the coordination motive (e.g., Cornand and Heinemann, 2014; (Shapiro et al., 2014)). Albeit partial transparency strategies are fruitfully discussed in the theoretical literature, they have yet received little attention by experimenters. Our work is a suite of the aforementioned papers, but it is mainly and closely related to Baeriswyl and Cornand (2014). Their experimental design implies a test of the “partially hidden information” strategy against a “less accurate public information” one. While both strategies share the ability to control overreaction to fully disclosed public information, the authors express a clear-cut preference for the “less accurate public information” strategy. Their argument essentially builds on feasibility and fairness issues when undertaking a partial transparency strategy in practice.

To the best of our knowledge, none of the previous experiments have tested a “fragmented information” strategy in laboratory. This gap in literature has basically motivated our current work. We study and compare the effects of two central bank's partial transparency strategies on subjects’ behavior in two separate experiments, namely the “fragmented information” and “partially hidden information” strategies. In a baseline and common treatment to both experiments, the central bank fully communicates its information. Within each treatment and round of an experiment, the central bank's (semi)-public information in addition to a private signal and the history play will not only help an individual guess the (unknown) true state of fundamentals, but also form beliefs about what her group's opponents think about the real state of fundamentals.⁸

An extension of the study by Baeriswyl and Cornand (2014) consists of supplying an argument based on a welfare comparative analysis that exclusively covers partial transparency strategies, in addition to the feasibility/fairness argument discussed in their paper. In Section 2, we hypothesize that a group of subjects is more rewarded under a “fragmented information” strategy than another group of subjects who play a “partially hidden information” strategy when the equality of the weights assigned to the semi-public signal is approved under both strategies. Besides, we engage in a welfare/payoff analysis because one advantage of the “fragmented information” strategy is the lower dispersion across agents’ behavior compared to the “partially hidden information” counterpart, which makes the former strategy politically more appealing.

Before testing partial transparency strategies, a fundamental question is whether subjects overreact to public information in order to approve the transition from full to partial transparency strategies. If overreaction to the (semi-)public signals is valid, one has to wonder whether subjects play the Nash equilibrium. Since perfect rationality and accurate beliefs about opponents’ strategies are hardly satisfied especially in the early periods of a game, subjects (or the groups of subjects) might diverge from the Nash equilibrium. Besides, subjects do not necessarily act in the same way. The distribution of the weights assigned on the (semi-)public signal, therefore, emphasizes heterogeneity of behavior that might be entailed by boundedly rational thinking (Nagel, 1995). Nevertheless, with multiple-round interactions and feedback, players are expected to revise their choices with experience and update their beliefs about their opponents’ strategies. Hence, another extension to the paper of Baeriswyl and Cornand (2014) would be to provide a deeper discussion on how learning process improves players’ behavior over time, which complements the role of the payoff incentives scrutinized in Baeriswyl and Conrnand (2014). Using aggregate data, we provide evidence of learning across treatments of a game. To get more accurate conclusions, we go through an individual analysis and appeal to the “choice reinforcement learning” (see Erev and Roth, 1998) and “belief-based learning” models (see (Fudenberg et al., 1998); Camerer and Ho, 1999; Feltovich, 2000; Camerer, 2003).⁹ These learning models pertain to “experiential” and “observational” learning approaches, respectively. Using panel regressions, we show that fitting data with one or the other learning model depends on the experiment in which subjects take part.¹⁰

The remainder of the paper is as follows. Section 2 describes the theoretical foundations; Section 3 presents the experimental environment and states hypotheses; we report and discuss our main results in Section 4; and Section 5 concludes with policy recommendations and practical limitations.