Job Market Paper

Dynamic Optimal Portfolio Construction with Reinforcement Learning

With Abdallah Aaraba (Department of Computer Science, University of Sherbrooke)

We adopt deep neural networks and reinforcement learning towards dynamic bottom-up portfolio construction by directly optimizing the weight of each asset conditioning on a large set of stock characteristics and macroeconomic indicators. Our model offers flexibility in the dynamics of returns, covariances, and trading costs, and allows explicit constraints on volatility and liquidity risks. In out-of-sample tests, our model substantially outperforms existing benchmarks, while keeping the transaction costs and turnover at the admissible minimum from real investment perspective. Our results are not driven by extensive short-selling or leverage, and remain robust under various economic restrictions and market conditions. In addition, we identify stock and macroeconomic features that are most important in building an economically feasible portfolio with forward-looking insights.

Presented at: McGill University, Wolfe Research Virtual Event (invited), Blackstone (invited)

Working Paper

Multi-(Horizon) Factor Investing with AI

With Ruslan Goyenko (McGill University)

We provide a novel approach for multi-factor investing with big data by a multi-horizon investor who takes into consideration long-term versus short-term volatility, liquidity and trading costs trade offs while maximizing expected portfolio returns. Reinforcement learning (RL), which is generally used to solve problems with long- versus short-term reward trade-offs, allows explicitly incorporating long, ten-year investment horizon considerations during training. In out-of-sample, testing period we are the first to show the importance of investment horizon effect for portfolio performance. First, RL portfolio of long term investors with annual rebalancing performs competitively vis-à-vis their short-term peers with monthly rebalancing, and outperforms the latter due to lower portfolio rebalancing needs, turnover and trading costs. Second, when both, short and long-term investors are allowed to rebalance monthly, long-horizon portfolio outperforms by being more patient, with more strategic factor timing and turnover strategies spread over multiple months. Short horizon portfolio is less patient, has higher volatility and almost twice higher monthly turnover. Importantly, we identify different fundamental economic signals determining success of long vs. short-term strategies.

Presented at: McGill University, CREDIT 2022 Conference: Long Run Risks, JP Morgan Asset Management Global Quant Council (invited), 2022 Global AI Finance Research Conference

Asset Pricing with Attention Guided Deep Learning

With Philippe Chatigny (University of Sherbrooke) and Ruslan Goyenko (McGill University)

Deep learning methods, which can accommodate wide ranges of various stock characteristics to identify optimal investment portfolio or stochastic discount factor (SDF), have been criticised for extracting their superior performances from difficult to arbitrage stocks, high limits-to-arbitrage market conditions or extreme turnovers. We introduce attention-guided deep learning, which allows, in a data driven way, identifying the most influential time-varying firm characteristics contributing to SDF. The attention guided SDF outperforms existing models after excluding small and micro-cap stocks, avoids extreme portfolio weights, turnovers and unlike other models, its performance is not driven by specific market regimes.

Presented at: University of Manitoba, Yale School of Management, 2022 INQUIRE Spring Residential Seminar, 4th Future of Financial Information Conference, McGill University Centre for Intelligent Machines, 2022 China International Conference in Finance

The Joint Cross Section of Option and Stock Returns Predictability with Big Data and Machine Learning

Reject & Resubmit, Review of Finance

With Ruslan Goyenko (McGill University)

Which market has leading informational advantage: stocks or options? Using large set of stock and option characteristics, and machine learning, we provide a comprehensive analysis of which characteristics are the first order importance predictors of option and stock returns. First, we find that option, rather than stock, characteristics are dominant predictors of option returns. Second, option, rather than stock, characteristics are also dominant predictors of stock returns. Consistent with the argument that an increase in trading activity in derivatives decreases information asymmetry about the underlying, option illiquidity is identified as the most important predictor of both stock and option returns.

Presented at: ITAM, 14th Financial Risks International Forum, Tilburg University, Virtual Derivatives Workshop, SoFiE Annual Conference

Liquidity Guided Machine Learning: The Case of the Volatility Risk Premium

With Eric Ghysels (University of North Carolina at Chapel Hill) and Ruslan Goyenko (McGill University

The financial industry has eagerly adopted machine learning algorithms to improve on traditional predictive models. In this paper we caution against blindly applying such techniques. We compare forecasting ability of machine learning methods in evaluating future payoffs on synthetic variance swaps. Standard machine learning methods tend to identify contracts which are illiquid, and hard to trade. The most successful strategies turn out to be those where we pair machine learning with institutional and market/traders inputs and insights. We show that liquidity guided pre-selection of inputs to machine learning results in trading strategies with improved pay-offs to the writers of variance swap contract replicating portfolio.

Demand Pressures and Option Returns

With Ruslan Goyenko (McGill University)

Delta-hedged option and straddle returns of S&P500 Index and equity options computed using end-of-day (EOD) closing prices are always higher compared to those based on any other price of the day. The difference between these returns can easily reach more than 100 bps per day or week. Options end-users’ demand pressures contribute to deviation of EOD prices from fundamental values. Computing returns using first half of the day prices, which are less distorted by demand pressures, helps explain several anomalies in the literature and establish identical volatility pricing across equity and index options.

Presented at: Canadian Derivatives Institute Conference, Virtual Derivatives Workshop

Work In Progress

Intra-day Option Returns: a Tail of Two Momentum

With Zhi Da (University of Notre Dame) and Ruslan Goyenko (McGill University)

Deep Learning for Financial Statements: Reading between the Lines using AI

With Nicolas Chapados (ServiceNow), Zhenzhen Fan (University of Manitoba), Ruslan Goyenko (McGill University), Issam Hadj Laradji (ServiceNow), and Fred Liu (University of Guelph)