Deep Partial Least Squares for Factor Modeling
Abstract
Across hedge funds, asset management, and proprietary trading firms, it is commonplace to use supervised learning for asset allocations and trade signal generation by performing ``feature engineering''. This often leads to a high dimensional input space. We present a high dimensional data reduction technique which uses partial least squares within deep learning. This framework provides a nonlinear extension of PLS together with a disciplined approach to feature selection and architecture design in deep learning. Unlike purely deep learning based data reduction techniques, such as autoencoders, we achieve the best of both worlds: fast and scalable SVD based algorithms for orthogonal projection combined with more parsimonious, yet highly expressive, deep architectures. Using 3290 assets in the Russell 1000 index over a period of December 1989 to January 2018, we assess a 49 factor model and generate information ratios that are approximately 50% greater than the OLS factor models and around 20-25% greater than deep learning. Furthermore, we observe that DPLS exhibits superior performance and robustness to outliers compared to OLS and deep learning. This is joint work with Nick Polson (Chicago Booth).
In this talk, we develop a RL approach to goal-based wealth management problems such as optimisation of retirement plans or target-dated funds. We present G-Learner: a generative RL algorithm that is suitable for noisy high dimensional data. In addition to quadratic regulators for G-Learners, which solve the direct RL problem very computationally efficiently, we develop GIRL, a G-learning inverse RL algorithm (GIRL) to infer the investor reward function from the observed trading actions. Examples demonstrating our G-learner and GIRL on high-dimensional historical data are presented. This is joint work with Igor Halperin (Fidelity Investments).
Bio
Matthew Dixon, Ph.D, FRM, began his career in structured credit trading at Lehman Brothers. He has consulted for numerous investment management, trading and financial technology firms in machine learning and risk analytics. His research focuses on mathematical algorithms for prediction, outlier detection, and risk, applying concepts in computational and applied mathematics to industrial modeling, especially in the area of investment management, algorithmic trading, and derivatives. He is the co-author of the 2020 textbook "Machine Learning in Finance: From Theory to Practice" and has written over 40 peer reviewed papers on machine learning, the blockchain, and quantitative finance, is RISK Magazine's Buy-side Quant of the Year (2022), the recipient of an Illinois Tech innovation award and the College of Computing's Dean Award for Excellence in Research (Junior level). He has been PI/co-PI on research funding from Intel, Dell, NASA JPL, and the NSF in addition to being quoted in the Financial Times and Bloomberg Markets. Matthew has recently co-authored the CFA course material on machine learning, serves on the CFA advisory committee for quantitative trading, and is associate editor of the World Scientific Annual Review of Fintech. He holds a PhD in Applied Math from Imperial College and has held visiting academic appointments at Stanford and UC Davis. Most recently, he founded a venture capital backed global stablecoin settlement network startup in Chicago.
