I am an assistant professor of finance at Columbia Business School. I currently work in the areas of asset pricing, asset management, and machine learning.
(with S. Abis)
Revise and Resubmit, Journal of Financial Economics
Conferences: AFA, NBER, FutFinInfo
Mutual fund prospectuses contain a wealth of qualitative information about fund strategies, yet a systematic analysis of this content is missing from the literature. We use machine learning to group together funds with similar strategy descriptions, and ask whether they act in accordance with the text. Despite weak legal recourse for investors, we find that mutual funds largely do keep their promises. We document a market-based disciplinary mechanism: when funds diverge from their group's core strategy, investors withdraw capital. Funds respond to these punitive outflows by reducing their divergence from the peer group average at a faster rate.
(with R. Di Mascio and N. Y. Naik)
Revise and Resubmit, Review of Financial Studies
Conferences: AFA, EFA, SFS Cavalcade, Jackson Hole Finance Conference, Copenhagen FRIC Conference, Luxembourg Asset Management Summit, AFFI/EUROFIDAI Paris December Meeting
We document novel facts about the term structure of institutional trading and performance using transaction-level data on professional fund managers. New stock purchases earn positive risk-adjusted returns that decay gradually over the subsequent twelve months, and managers continue to buy the same stock in small increments for as long as the alpha remains positive, with proportional intensity. Greater competition for information and more highly correlated signals are associated with more aggressive trading and lower alpha. Our findings confirm many predictions of informed trading models, but also pose some new challenges for the theoretical literature.
(with S. Ke)
We use deep Bayesian neural networks to investigate the determinants of trading activity in a large sample of institutional equity portfolios. Our methodology allows us to evaluate hundreds of potentially relevant explanatory variables, estimate arbitrary nonlinear interactions among them, and aggregate them into interpretable categories. Deep learning models predict trading decisions with up to 86% accuracy out-of-sample, with market liquidity and macroeconomic conditions together accounting for most (66-91%) of the explained variance. Stock fundamentals, firm-specific corporate news, and analyst forecasts have comparatively low explanatory power. Our results suggest that market microstructure considerations and macroeconomic risk are the most crucial factors in understanding financial trading patterns.
(with S. Abis, A. M. Buffa, and A. Javadekar)
Conferences: AFA, SFS Cavalcade, FutFinInfo, UNSW AP Workshop, Melbourne AP Meeting, TAU Finance Conference
We analyze fund managers' incentives to disclose qualitative information about their strategies, and investors’ ability to learn from these disclosures. We propose a mechanism whereby investors make fewer errors in distinguishing active returns from passive factor exposures when they have access to more detailed strategy descriptions. In a formal model, we show that investor attribution errors are, on balance, more costly for managers with more specialized strategies, leading them to write more detailed descriptions. In the data, we find evidence for this prediction and support for the model’s core learning mechanism, as well as new insights into the flow-performance relationship.
Conferences: NBER, EFA
I show that fund managers who are compensated for relative performance optimally shift their portfolio weights towards those of the benchmark when volatility rises, putting downward price pressure on overweight stocks and upward pressure on underweight stocks. In quarters when volatility rises most (top quintile), a portfolio of aggregate-underweight minus aggregate-overweight stocks returns 2% to 5% per quarter depending on the risk adjustment. Placebo tests on institutions without direct benchmarking incentives show no effect. My findings cannot be explained by fund flows and thus constitute a new channel for the price effects of institutional demand.
Work in Progress
Reinforcement Learning in Asset Pricing
Reinforcement learning (RL) algorithms can be used to efficiently solve complex discrete time economic systems that are computationally too expensive for standard numerical methods. I introduce a Walrasian auctioneer into the popular Actor-Critic family of RL algorithms to allow for market clearing, and apply this new methodology to solve a dynamic equilibrium model of asset pricing under asymmetric information. The model features many assets with an arbitrary covariance structure, multiple strategic investors with heterogeneous private signals, uninformed non-strategic investors, and transaction costs. Unlike in standard strategic trading models, informed trading intensity in my model is reduced when the fraction of informed traders in the market rises, while return volatility is increased. The model generates complex trading dynamics, where investors with more precise private signals learn to front-run investors with less precise signals, leading to price overreactions and corrections despite all agents having rational expectations.
Trade-Based Performance Measurement
(with R. Di Mascio and N. Naik)
We propose new metrics for investment performance based on short-run trading profitability. Since investment opportunities are scarce and value-relevant information decays over time, marginal decisions made by fund managers (i.e., trades) should provide more accurate signals about underlying skill than portfolio alphas, which are contaminated by the returns on "stale" positions. Our measures range from the very simple ("hit rate", or the fraction of trades that outperform the benchmark over the subsequent month) to the more complex (regressions relating trade size to subsequent profitability). We examine the validity of these measures in a global sample of long-only equity funds, for which we observe daily trading activity. In our sample, trade-based metrics are more persistent than portfolio alphas and, more importantly, are better able to forecast future portfolio alphas (in a mean squared error sense). Simple and complex methods are almost equally effective. A hypothetical manager-selection exercise reveals that trade-based performance measurement can improve the risk-adjusted returns to investors by up to 3% per annum.
A Macro-Finance Model of Carbon Pricing (with N. Clara)
Active and Passive Management: A Unified Approach (with P. Akey)