The Informational Content of Mutual Fund Prospectuses

Abstract: The drafting of prospectuses entails significant costs for mutual funds, yet the SEC has not demonstrated that they are actually relevant for investors. We use machine learning to study the importance of the textual information contained in prospectuses of a broad sample of US active equity mutual funds, available via EDGAR (the SEC’s online reporting system). Using supervised learning on ex-ante prospectuses we are able to predict which funds are more likely to engage in destructive, agency-related risk-shifting behavior up to 3 years ahead. We also use unsupervised learning to group funds into distinct clusters based on the similarity in prospectus text, uncovering groups with natural interpretations such as “quantitative”, “macro-focused”, “sector-timing”, “regulatory/political risk”, “derivatives risk”, etc. We find that membership in particular clusters is predictive of the shape of funds’ future return distributions.

Bio: Professor Simona Abis joined Columbia Business School in 2017. She holds a PhD from INSEAD. Before joining the PhD program Simona worked as a quantitative researcher for a systematic hedge fund. Her research interests span the fields of information economics, empirical and theoretical asset pricing, machine learning, mutual funds and hedge funds. Overall Simona is interested in the impact of technology on financial markets. Her current research focuses particularly on the impact of technological change on investment management through the rise of quantitative investment and on identifying the informational content of funds’ mandatory disclosures to the regulator.

View slides

Learning strategic interaction from individual action: A game-theoretic approach

Abstract:

Individuals, or organizations, cooperate with or compete against one another in a wide range of practical situations. In the economics literature, such strategic interaction are often modeled as games played on networks, where an individual’s payoff depends not only on her action but also that of her neighbors. The current literature has largely focused on analyzing the characteristics of network games in the scenario where the structure of the network, which is represented by a graph, is known beforehand. It is often the case, however, that the actions of the players are readily observable while the underlying interaction network remains hidden. In this paper, we propose two novel frameworks for learning, from the observations on individual actions, network games with linear-quadratic payoffs, and in particular the structure of the interaction network. Our frameworks are based on the Nash equilibrium of such games and involve solving a joint optimization problem for the graph structure and the individual marginal benefits. We test the proposed frameworks in synthetic settings and further study several factors that affect their learning performance. Moreover, with experiments on three real world examples, we show that our methods can effectively and more accurately learn the games than the baselines. The proposed approach is among the first of its kind for learning quadratic games, and have both theoretical and practical implications for understanding strategic interaction in a network environment.

Bio:

Yan Leng is a Ph.D. candidate at the MIT Media Lab and a Research Assistant in the Human Dynamics Group, led by Professor Alex Pentland. She received dual masters from MIT in Computer Science and Transportation Engineering in 2016. Her research interests lie at the intersection of machine learning, statistical inference, and social sciences. She is passionate about human behaviors, understanding how they make decisions in social networks, and modeling their behaviors with computational methods. In particular, she develops graph-based machine learning techniques for network inference, behavior prediction, and counterfactual prediction for causal inference.

View slides

How to Play Fantasy Sports Strategically (and Win)

Abstract:

Daily Fantasy Sports (DFS) is a multi-billion dollar industry with millions of annual users and widespread appeal among sports fans across a broad range of popular sports. Building on the recent work of Hunter, Vielma and Zaman (2016), we provide a coherent framework for constructing DFS portfolios where we explicitly model the behavior of other DFS players. We formulate an optimization problem that accurately describes the DFS problem for a risk-neutral decision-maker in both double-up and top-heavy payoff settings. Our formulation maximizes the expected reward subject to feasibility constraints and we relate this formulation to the literature on mean-variance optimization and the out-performance of stochastic benchmarks. Using this connection, we show how the problem can be reduced to the problem of solving a series of binary quadratic programs. We also propose an algorithm for solving the problem where the decision-maker can submit multiple entries to the DFS contest. This algorithm is motivated by some new results for parimutuel betting which can be viewed as a special case of a DFS contest. One of the contributions of our work is the introduction of a Dirichlet-multinomial data generating process for modeling opponents’ team selections and we estimate the parameters of this model via Dirichlet regressions. A further benefit to modeling opponents’ team selections is that it enables us to estimate the value in a DFS setting of (i) insider trading and (ii) collusion whereby a number of DFS players combine to construct a single portfolio of entries to a given contest. We demonstrate the value of our framework by applying it to both double-up and top-heavy DFS contests during the 2017 NFL season.

Bio:

Martin Haugh is an Associate Professor of Analytics and Operations Research at Imperial College, London. Before joining Imperial in 2017 he spent over 10 years in the IEOR department at Columbia University and 4 years working as a quant in the hedge-fund industry. His research interests include finance/risk management, data science and dynamic programming/stochastic control. He earned his PhD in Operations Research from MIT in 2001.

View slides

Using AI Machine Learning to Explore Large Streaming Financial Data Sets to Improve Market Making

Abstract: The dollar treasury and futures market is arguably the most liquid and actively traded market in the world. On a daily basis, this market generates hundreds of millions of records of data, and the set of securities involved amounts to thousands. The environment is dynamic, interrelated, and fast-paced, and liquidity conditions are constantly changing. As participants engage in strategies, enter and exit this marketplace, certain relationships capture the attention of various participants, who apply capital to their (sometimes machine-learned) convictions. Moreover, correlations are non-static and exhibit a term structure. Instead of making a single static model, we will explore how using Multi-GPU setups and this large streaming dataset, you can set up an online machine learning environment where thousands of strategies can be monitored and pockets of available liquidity uncovered.

Bio: Peter works in the Rates Trading group at Citi. He focuses on machine learning for the implementation of pricing and risk analytics. Peter has developed neural net applications for natural language processing, as well as probabilistic graphical models for pricing. Peter joined Citibank’s Fixed Income Algo trading group in 2011. This team has deployed the largest bank systematic trading and execution platform for treasuries and bond futures.   

Prior to joining Citi, Peter headed the Rates Group at Quantifi. There, he was responsible for managing the product development process for all Rates, Convertible Bonds and FX Options Solutions within the Quantifi product suite. Peter started his career in Research and Technology at Bear Stearns before heading fixed income derivatives research for Deutsche Bank. He has traded fixed income derivatives (linear and non-linear products), government bonds and agencies for Lehman Brothers and Salomon Brothers. He headed the fixed income derivatives trading desk for a number of European banks. Peter has worked on GPU-based software applications in his areas of expertise for more than ten years. He has appeared as a Speaker on the use of GPUs for the computation of risk, pricing of exotic derivatives and HPC in general, at Nvidia and Microsoft conferences.

Redefining NYC neighborhoods using open data and machine learning

Abstract: New York City’s (NYC’s) neighborhoods are a driving force in the lives of New Yorkers—their identities are closely intertwined and a source of pride. However, the history and evolution of NYC’s neighborhoods don’t follow the rigid, cold lines of statistical and administrative boundaries. Instead, the neighborhoods we live and work in are the result of a more organic confluence of factors. NewerHoods is an interactive web-app that uses open data to generate localized features at the US Census tract-level then clusters them spatially to define a data-driven neighborhood. Traditional clustering algorithms tend to be a-spatial or indifferent to the geographic adjacency of tracts which is problematic when identifying neighborhoods. To solve this, we use a spatial hierarchical clustering model which allow us to balance between spatial vicinity and similarity in the feature space. Users are able to select characteristics of interest (currently open data on housing, crime, and 311 complaints), visualize NewerHood clusters on an interactive map, find similar neighborhoods, and compare them against existing administrative boundaries.  The tool is designed to enable users without in-depth data expertise to compare and incorporate these redefined neighborhoods into their work and life. This presentation will dig into the methodology behind NewerHoods.

NewerHoods was developed by the Two Sigma Data Clinic. The Two Sigma Data Clinic develops pro bono solutions that enable social impact organizations to use data and technology more effectively, and have a greater impact on the communities they serve. As data-driven decision-making has proliferated across sectors, nonprofits have lagged behind due to funding and resource constraints. While they may be in the early stages of data collection, the widespread availability of open data has the potential to fill these gaps and inform nonprofits’ operations and programming.

Bio: Darren Vengroff is a Senior Vice President at Two Sigma, where he leads the AI Engineering team. Previously, Darren served as Chief Scientist at two startups (Rich Relevance and Meld), and CTO of a third (Pelago). Prior to that he was a Principal Engineer at Amazon.com where he developed personalization systems, a Core Strategist at Goldman Sachs and a Software Engineer at Microsoft. He has also served as an advisor to the Gates Foundation and a number of startups. Darren holds both an M.S. and a Ph.D. in Computer Science from Brown University, as well as a B.S.E. in Computer Science from Princeton University

View slides

Text Analytics in Finance

Abstract: Text can be a great source of market signals. In the past, most text analytics for market signal extraction was fairly simple, word count based. Now, as finance firms build data science groups, the complexity and scale of the available text analytics is much greater. However, the basic central problems remain - acquire text data, clean text data, link text data to market-relevant entities and events. In this talk, I will showcase some applications of text analytics in finance, and some differences between text analytics applications in the financial industry and "standard" NLP.

Bio: Amanda Stent is a NLP architect at Bloomberg. Previously, she was a director of research and principal research scientist at Yahoo Labs, a principal member of technical staff at AT&T Labs - Research, and an associate professor in the Computer Science Department at Stony Brook University. Her research interests center on natural language processing and its applications, in particular topics related to text analytics, discourse, dialog and natural language generation. She holds a PhD in computer science from the University of Rochester. She is co-editor of the book Natural Language Generation in Interactive Systems (Cambridge University Press), has authored over 90 papers on natural language processing and is co-inventor on over twenty patents and patent applications. She is president emeritus of the ACL/ISCA Special Interest Group on Discourse and Dialog, treasurer of the ACL Special Interest Group on Natural Language Generation and one of the rotating editors of the journal Dialogue & Discourse. She is also a board member of CRA-W, where she co-edits the newsletter.

View slides

Forecasting price moves from order flow: perspectives from Deep Learning

Abstract: Using a large-scale Deep Learning approach applied to a high-frequency database containing billions of electronic market quotes and transactions for US equities, we uncover nonparametric evidence for the existence of a universal and stationary price formation mechanism relating the dynamics of supply and demand for a stock, as revealed through the order book, to subsequent variations in its market price. We assess the model by testing its out-of-sample predictions for the direction of price moves given the history of price and order flow, across a wide range of stocks and time periods. The universal price formation model is shown to exhibit a remarkably stable out-of-sample prediction accuracy across time, for a wide range of stocks from different sectors. Interestingly, these results also hold for stocks which are not part of the training sample, showing that the relations captured by the model are universal and not asset-specific.

Joint work with Justin Sirignano (Urbana Champaign)

Based on:
Universal Features of Price Formation in Financial Markets: Perspectives From Deep Learning 
https://ssrn.com/abstract=3141294

View slides