Lu Hong’s research while affiliated with Loyola University Chicago and other places

Many decisions and actions can be framed as binary classification problems in which an outcome function maps states of the world into one of two outcomes and in which individuals use models (interpreted signals) to classify the state. For this class of binary classification problems, we fully characterize the range of possible group accuracies as a function of group size, average individual accuracy and diversity (average pairwise disagreement) or groups using majority rule. Our characterization yields five implications (i) the range of possible collective accuracies can be large, especially for groups of low accuracy individuals, (ii) up to moderate levels of diversity, the maximal collective accuracy gain equals the maximal collective accuracy loss, (iii) possible group accuracy set-wise improves in the average accuracy of their members, (iv) larger groups increase the range of possible collective accuracies unless diversity is high, and (v) for groups to be guaranteed to be more accurate than their average member, diversity must be high.

In this paper, we derive necessary and sufficient conditions on team based tasks in order for a selection criterion applied to individuals to produce optimal teams. We assume only that individuals have types and that a team’s performance depends on its size and the type composition of its members. We first derive the selection principle which states that if a selection criterion exists, it must rank types by homogeneous team performance, the performance of a team consisting only of that type. We then prove that a selection criterion exists if and only if replacing the team’s lowest ranked type, as measured by homogeneous team performance, with a higher ranked type increases team performance. Finally, we show that the replace the lowest ranked property rules out most common types of team complementarities, including benefits to diverse types and types that fill structural holes.

An increasing proportion of decisions, design choices, and predictions are being made by hybrid groups consisting of humans and artificial intelligence (AI). In this paper, we provide analytic foundations that explain the potential benefits of hybrid groups on predictive tasks, the primary use of AI. Our analysis relies on interpretive and generative signal frameworks as well as a distinction between the big data used by AI and the thick, often narrative data used by humans. We derive several conditions on accuracy and correlation necessary for humans to remain in the loop. We conclude that human adaptability along with the potential for atypical cases that mislead AI will likely mean that humans always add value on predictive tasks.

Markets, democracies, and organizations rely on accurate aggregate predictions to function properly. A large literature explains how accuracy can arise from diverse predictive models, typically captured as independent or non perfectly correlated signals. Yet, that literature largely ignores how the diversity of models arises and is maintained. In this paper, we derive equilibrium levels of model diversity as a function of social structure, population size, the probability of experimentation, and the number of available models by building on a theoretical framework used to study biodiversity. We then link model diversity to collective accuracy by generalizing the bias-variance decomposition formula. Assuming equally accurate models, we find that for large populations collective accuracy depends primarily on the diversity of available models and that for small populations, social structure and rates of experimentation also matter. We then show, contrary to intuition, that dividing a population into isolated sub groups does little to increase equilibrium diversity levels. We also extend the model to allow for heterogeneity in accuracy and selection effects.

In this article, we develop a novel model to forecast the volatility of S&P 500 futures returns by considering measures of limits to arbitrage. When arbitrageurs face constraints on their trading strategies, option prices can become disconnected from fundamentals, resulting in a distortion that reflects the limits to arbitrage. The corresponding market based implied volatility will therefore also contain these distortions. Our contributions are both conceptual and empirical. Conceptually, the limits to arbitrage framework can shed light on relative asset prices as exemplified by this particular study. Empirically, our volatility forecasting model explains 71% of the variation in realized volatility, a substantial improvement over a naive forecast based only on lagged realized volatility, which produces an R2 of 53%. © 2014 Wiley Periodicals, Inc. Jrl Fut Mark

In this article, we develop a novel model to forecast the volatility of S&P 500 futures returns by considering measures of limits to arbitrage. When arbitrageurs face constraints on their trading strategies, option prices can become disconnected from fundamentals, resulting in a distortion that reflects the limits to arbitrage. The corresponding market based implied volatility will therefore also contain these distortions. Our contributions are both conceptual and empirical. Conceptually, the limits to arbitrage framework can shed light on relative asset prices as exemplified by this particular study. Empirically, our volatility forecasting model explains 71% of the variation in realized volatility, a substantial improvement over a naive forecast based only on lagged realized volatility, which produces an R-super-2 of 53%. © 2014 Wiley Periodicals, Inc. Jrl Fut Mark 35:987–1002, 2015

In many applications, agents (whether human or computational) provide estimates that must be combined at a higher level. Recent research distinguishes two kinds of such estimates: interpreted and generated data. These two kinds of data require different kinds of aggregation processes, which behave differently from an information geometric perspective: interpreted estimates require methods such as voting that can leave the convex hull of the individual estimates, while the optimal aggregation for generated estimates lies within the convex hull and thus is accessible by methods such as weighted averages. We motivate our analysis in the context of a crowdsourced forecasting application, demonstrate the central insights theoretically, and show how these insights manifest them-selves in actual data.

In this paper, we construct a framework within which we explore how incentives and information structures influence the ability of a collection of individuals to make an accurate aggregate prediction. In our framework, individuals of bounded ability predict outcomes that depend on the values of a set of attributes. Individual construct models consider only a subset of those attributes, and those models depend on their incentives and their information environments. We consider two types of incentive structures: one in which individuals get paid on the basis of accuracy and one based on market like, for example, parimutuel payoffs. We also consider two information environments: one in which individuals learn in isolation and another in which they can copy more successful predictors. We find that market incentives and isolated learning environments produce the most accurate aggregate predictions but that these same incentives and information structures also produce the least accurate individuals. Thus, the incentives and informational structures that produce collective wisdom may hinge on their ability to produce and maintain diversity.

Traditional economic models typically treat private information, or signals, as generated from some underlying state. Recent work has explicated alternative models, where signals correspond to interpretations of available information. We show that the difference between these formulations can be sharply cast in terms of causal dependence structure, and employ graphical models to illustrate the distinguishing characteristics. The graphical representation supports inferences about signal patterns in the interpreted framework, and suggests how results based on the generated model can be extended to more general situations. Specific insights about bidding games in classical auction mechanisms derive from qualitative graphical models.