Reading Group: Machine Learning for Asset Managers

In this lecture, we'll be discussing how machine learning (ML) techniques can be used to analyze complex financial data. We'll explore the benefits of using ML in financial research, and how it compares to traditional econometric approaches. We'll also delve into the challenges of working with unstructured and high-dimensional data, and how ML can help overcome these challenges. Whether you're an econometrician looking to incorporate ML into your work or simply want to learn more about the role of technology in finance, this video is for you.

This lecture explores the power of deep learning in predicting high-frequency returns. By training artificial neural networks on stationary inputs derived from the order book, the study achieves state-of-the-art predictive accuracy, surpassing traditional models.

Our lecture navigates into asset pricing with machine learning unveiled a nuanced dance between simplicity and complexity.

While Occam's Razor advocates for simplicity, our study revealed the dominance of intricate models like Neural Networks and Random Forests over traditional linear ones like OLS in predicting asset returns. Notably, Neural Network 3 (NN3), with its modest three hidden layers, consistently outshines the rest, showcasing the power of simplicity in unraveling complex financial data.

Furthermore, machine learning's prowess extends to large-cap stocks, challenging conventional market wisdom.

We examine core algorithms such as Neural Networks, DNN, Fuzzy Logic, and SVMs to uncover a holistic view of intricate variations in these foundational techniques.

This lecture examines a new AI-driven Peer Grouping System (AIPGS) that can group stocks according to market perception. AIPGS groups together companies with similar fundamental characteristics, resulting in a higher level of diversification potential for portfolios constructed using ML.

This lecture examines an approach for identifying lead-lag clusters within multivariate systems, treating pairwise relationships as a directed network.

Model Fingerprint algorithm decomposes feature impact on linear, non-linear, and pairwise interaction effects. The algorithm was described in Li, Turkington, Yazdani "Beyond the Black Box: An Intuitive Approach to Investment Prediction with Machine Learning" Shapley Feature Importance is a game-theory approach applied to estimate the importance of a feature.

The lecture explores the journey of the attention mechanism, which emerged from the roots of RNN and LSTM to its role in transformer networks. This ascent of the attention mechanism has been a pivotal force in shaping our strategies.

The lecture considers the potential of ChatGPT-4 to forecast stock market movements, contrasting its capabilities with previous iterations like GPT-1, GPT-2, and the BERT model. We examine the model's advancements but also the pros and cons of deploying ChatGPT-4 in trading algorithms, addressing the risks involved.

The lecture examines various aspects of the model, including its architecture and origins, the distinctive dataset used to train it, its assessment for financial tasks, and the extent to which it lives up to the media's hype regarding its relevance and groundbreaking nature.

The lecture discusses how meta-labeling is a machine learning (ML) layer that sits on top of a base primary strategy to help size positions, filter out false-positive signals, and improve metrics such as the Sharpe ratio and maximum drawdown.

The lecture will consider separating the side and size of a position allows for sophisticated strategy structures to be developed. Modeling the size component can be done through a meta-labeling approach. This lecture establishes several heterogeneous architectures to account for key aspects of meta-labeling. They serve as a guide for practitioners in the model development process, as well as for researchers to further build on these ideas.

Meta-labeling is a technique first introduced by Dr. Marcos Lopez de Prado, which can be used to solve the problem of non-stationarity as well as provide position sizes. The core idea is to increase position sizes in the likelihood of a positive event and reduce the size as uncertainty increases.

This lecture considers different ensemble methods for meta-labeling in finance and presents a framework to facilitate the selection of ensemble learning models for this purpose. Experiments were conducted on the components of information advantage and modeling for false positives to discover whether ensembles were better at extracting and detecting regimes and whether they increased model efficiency. We demonstrate that ensembles are especially beneficial when the underlying data consists of multiple regimes and is non-linear. Our framework serves as a starting point for further research. We suggest that the use of different fusion strategies may foster model selection. Finally, we elaborate on how additional applications, such as position sizing, may benefit from our framework.

This lecture discusses strategies to incorporate transaction costs effectively, manage portfolio constraints, and refine our approach to the sensitive estimates of expected returns and covariances.

This lecture explores the innovative FinRL library, which aims to democratise quantitative finance and stock trading strategy development, especially for beginners. The library is built upon deep reinforcement learning (DRL) techniques and consists of three layers: market environments, agents, and applications. It provides a comprehensive suite of tools, including diverse stock market datasets, neural network-based agent training, and robust backtesting capabilities. Moreover, FinRL incorporates crucial trading constraints like transaction costs, market liquidity, and risk aversion.

This lecture explores how DRL can revolutionize decision-making in finance. This paper bridges the gap between traditional asset management techniques and deep reinforcement learning (DRL) by demonstrating how DRL can shed new light on portfolio allocation. By treating portfolio allocation as a continuous control optimization problem with delayed rewards, DRL offers advantages such as adapting to changing market conditions, not relying on traditional risk assumptions, and incorporating additional data.

In our previous session, we examined the authors' proposed solution of utilizing deep reinforcement learning to address the longstanding challenge of portfolio allocation. This lecture, we will delve even deeper into the topic and expand our discussion to encompass the authors' further research. Notable contributions include expanding the scope of assets considered, seamlessly incorporating additional market features, leveraging supervised methods within the pipeline, and introducing a host of other intriguing additions.

The focus is on understanding the Diversification Ratio, the Most Diversified Portfolio (MDP), and its application in a mean-variance framework. The lecture also considers the Portfolio Invariance Properties essential for unbiased portfolio construction.

The lecture will examine backtest the strategies, and use XGBoost to regress the Calmar ratio spread. With Shapley values, we'll uncover key factors driving performance differences, focusing on drawdown measures.

This lecture will focus on the problem of the fact that market timing is challenging but diversifying into "alternatives" like private equity, real estate, credit, and liquid alternatives can offer valuable risk diversification tools, especially in unfavorable market conditions with potentially longer-lasting downturns.

This lecture discovers the world of graph theory and its pivotal role in portfolio optimization. Through innovative tools like the Minimum Spanning Tree (MST) and the Triangulated Maximally Filtered Graph (TMPG), we gain insights into asset centrality and connections. These insights serve as the foundation for strategic asset selection within portfolio optimization models.

In this lecture we will examine: Causality is proposed as an advancement to the field of factor investing using causal graphs to reshape the landscape of factor investing. The focus will be on crucial ideas about factors, causal relationships and their impact on investment strategies.

This lecture examines many claims about factor investing; some are timeless, while others are focused on specific concerns that have emerged recently.

How should the theory of factor investing be implemented in practice? In this reading group session, we take a look at an article called "Factor Investing: The best is yet to come" which attempts to answer this question by discussing the selection of factors, how to combine them, and portfolio construction. The article also highlights next-generation developments in the factor investing space.

This lecture provides an analysis of the value and momentum strategies in equity markets, and how they can be used to construct a diversified portfolio that outperforms the market.

This lecture discusses the beta anomaly and how it can be exploited through a betting against beta (BAB) strategy. It also explores the potential risks and limitations of this approach and share insights on how this research can be applied.

The lecture will consider a clustering-driven methodology that uncovers key dependencies and insights from lagged multi-factor models. It will examine how this approach can enhance control, forecasting, and clustering across various domains, including financial markets and environmental datasets.

This lecture examines the partitioning of the correlation matrix of market residual returns of stocks into distinct clusters. By employing various clustering methods, we will learn how to effectively group stocks based on the similarity of their market residual returns.

Within each identified cluster, we consider constructing and evaluating the performance of mean-reverting statistical arbitrage portfolios. This method capitalizes on temporary price inefficiencies, expecting that prices will eventually revert to their mean.

This lecture tackles the complexities of managing multiple pairs in statistical arbitrage. It introduces a novel method using preference relations to harmonize conflicting trade signals across various security pairs.

We are honored to introduce Soohun Kim, an Assistant Professor of Finance at KAIST, who will be presenting his paper, "Characteristic-based Returns: Alpha or Smart Beta?" co-authored by Robert A. Korajczyk and Andreas Neuhierl. This paper has already received significant recognition, having won a Special Distinction Award from the Journal of Investment Management in 2022.

The code for "Why You Should Hedge Beta and Sector Exposures (Part I)”: https://github.com/marketneutral/lectures/blob/master/Lecture-Why%2BHedge%2BI-.ipynb

The lecture considers the complexities behind common risk. Each trading algorithm may carry only a small amount of common risk, but when combined, these risks can unexpectedly compound, challenging the conventional wisdom of diversification. It examines the paradox that diversification, while beneficial, can sometimes lead to an unintended concentration of common risk.

This lecture examines an approach to investing across markets with the "Economic Trend" strategy. By capitalizing on trends in macroeconomic fundamentals, this strategy consistently delivers attractive risk-adjusted returns over a 50+ year sample. It complements price trend-following and provides robust drawdown protection, improving long-term performance and resilience during market stress.

The lecture considers how to use over a hundred features and five machine learning algorithms to achieve superior out-of-sample forecasts. The system's effectiveness remains consistent across different specifications and can be scaled using hyperparameter transfer learning.

This lecture investigates the paradoxes of penny-picking. It uncovers how seemingly irrational behavior could be a strategic play over time, influenced by skewness-seeking and supported by classical behavioral theories.

This lecture establishes a framework for assessing how stop-loss rules affect both the expected return of a portfolio. In a market exhibiting momentum, these rules can actually add value.

This lecture explores how the fusion of mean reversion and momentum strategies can deliver potent risk-adjusted returns.

This lecture considers the impact of sustained inflation on the financial markets and how it affects the repositioning of investor portfolios.

This lecture examines risk-mitigating strategies and uncovers the strengths and weaknesses of regular index put buying ("Put") and multi-asset trend following ("Trend") as tail hedges.

This lecture uncovers essential statistical properties of asset returns, highlighting distributional, tail, and temporal dependencies. These findings pose challenges to common statistical approaches used in financial data analysis.