Research

Journal of Financial Economics, Volume 138, Issue 3, Pages 754-776 (December 2020). 

We consider an economy populated by investors with heterogeneous preferences and beliefs who receive non-pledgeable labor incomes. We study the effects of collateral constraints that require investors to maintain sufficient pledgeable capital to cover their liabilities. We show that these constraints inflate stock prices, give rise to clusters of stock return volatilities, and produce spikes and crashes in price-dividend ratios and volatilities. Furthermore, the mere possibility of a crisis significantly decreases interest rates and increases Sharpe ratios. The stock price has a large collateral premium over non-pledgeable incomes. Asset prices are in closed form, and investors survive in the long run. 

• European Finance Association Best Conference Paper Award, 2016

• SFS Finance Cavalcade Best Paper Award in Asset Pricing, 2016

• European Winter Finance Symposium Best Paper Prize in Memory of S. Bhattacharya, 2016

Revise and resubmit, Journal of Finance

This paper studies the dynamics of information acquisition and uncertainty in a noisy rational expectations model. Investors choose to acquire most information at times when uncertainty and risk premia are high; this choice feeds back and endogenously reduces subsequent uncertainty. Within the model, uncertainty is measured directly from risk-neutral variance---analogous to the VIX index---so this translates into the concrete prediction that risk-neutral variance mean-reverts rapidly following spikes in volatility, as is observed empirically. The cyclicality of information acquisition depends on the skewness of the underlying asset: if the market is negatively skewed, market-level information acquisition is countercyclical. Conversely, information acquisition and risk premia are high following good news for positively skewed assets such as individual stocks, which gives rise to momentum in the stock market. 

• Finance Theory Group Best Paper in Finance Theory on the Job Market, 2019

• Artificial intelligence (AI) increasingly interacts with human decision making, both by assisting users and by autonomously executing tasks. Because AI systems can learn from human actions, the composition of human users influences the evolution of the technology itself. This paper develops a parsimonious framework in which heterogeneous agents with innovative and executional skills decide whether to use AI and whether to remain active or passively delegate decisions to it. The model distinguishes between assistive AI, where humans retain final control, and autonomous AI, where the technology makes the final decision. Two forces arise as AI improves. A cleansing effect occurs when improved technology crowds out the weakest active users, raising the average quality of human decisions that can be used to train AI. A countervailing restriction effect emerges when the technology crowds out precisely the human skills most valuable for training. The model predicts that assistive AI attracts active users with stronger execution skills, whereas autonomous AI attracts users with stronger innovative skills, generating distinct patterns in who benefits from AI and how human participation shapes technological progress.

• The paper develops a model of bubbles that can be taken to the data and explain the behavior of asset prices and their statistics. We depart from the rational expectations framework and assume that investors are only boundedly rational. They observe the price process, but do not fully understand how its volatility and expected returns are determined in equilibrium. Investors learn about the market by looking at past prices. When they observe unexpectedly high returns, they infer that the asset must currently have a high Sharpe ratio, and therefore, allocate a higher share of their wealth to the asset, further increasing the asset price. The interaction of this feedback effect with investors' wealth effect determines the price dynamics and evolution of investors' beliefs in the model. We fit the model to cryptocurrency markets and show that it can successfully explain many empirical facts in these markets.

• We study the disagreement of major foreign exchange (FX) dealers using proprietary survey data on dealers' price quotes of short- and long-tenor currency derivatives. Dispersion among dealers is the highest at short tenors where heterogeneous information is of great relevance, and much lower at long tenors where heterogeneous beliefs dominate. This downward sloping term structure of dealer dispersion is most steep for risk reversals that capture asymmetric tail risk, and flattens considerably for forwards, strangles, and straddles that capture mean, symmetric tail risk, and volatility. Furthermore, dealer dispersion on risk reversals positively predicts currency returns in the cross section, with strong economic and statistical significance at short horizons but weak significance at long horizons. Dealer dispersion on the other three FX derivatives has no return predictive power. 

• Investors use a range of imperfectly observed predictor variables to form dividend expectations and update the sensitivities of dividend growth to these predictors. Investors are more certain about their prediction model after large predicted dividend changes. Significant dividend surprise, however, raises model uncertainty and volatility of future returns. Unpredicted comovement of stocks with similar characteristics increases these stocks’ risk premia. Information acquisition tends to focus on variables useful for short-term forecasts, allowing the buildup of model uncertainty for other predictors. These learning behaviors also generate correlated movements of expectation and subjective uncertainty of future dividend growth, inducing correlations in cash flow and discount rate components of asset returns.