Machine Learning for Factor Investing : Python Version

"Machine learning (ML) is progressively reshaping the fields of quantitative finance and algorithmic trading. ML tools are increasingly adopted by hedge funds and asset managers, notably for alpha signal generation and stocks selection. The technicality of the subject can make it hard for non-specialists to join the bandwagon, as the jargon and coding requirements may seem out-of-reach. Machine learning for factor investing: Python version bridges this gap. It provides a comprehensive tour of modern ML-based investment strategies that rely on firm characteristics. The book covers a wide array of subjects which range from economic rationales to rigorous portfolio back-testing and encompass both data processing and model interpretability. Common supervised learning algorithms such as tree models and neural networks are explained in the context of style investing and the reader can also dig into more complex techniques like autoencoder asset returns, Bayesian additive trees and causal models. All topics are illustrated with self-contained Python code samples and snippets that are applied to a large public dataset that contains over 90 predictors. The material, along with the content of the book, is available online so that readers can reproduce and enhance the examples at their convenience. If you have even a basic knowledge of quantitative finance, this combination of theoretical concepts and practical illustrations will help you learn quickly and deepen your financial and technical expertise"-- Provided by publisher I Introduction 1 Notations and data 1.1 Notations 1.2 Dataset 2 Introduction 2.1 Context 2.2 Portfolio construction: the workflow 2.3 Machine learning is no magic wand 3 Factor investing and asset pricing anomalies 3.1 Introduction 3.2 Detecting anomalies 3.2.1 Challenges 3.2.2 Simple portfolio sorts 3.2.3 Factors 3.2.4 Fama-MacBeth regressions 3.2.5 Factor competition 3.2.6 Advanced techniques 3.3 Factors or characteristics? 3.4 Hot topics: momentum, timing, and ESG 3.4.1 Factor momentum 3.4.2 Factor timing 3.4.3 The green factors 3.5 The links with machine learning 3.5.1 Short list of recent references 3.5.2 Explicit connections with asset pricing models 3.6 Coding exercises 4 Data preprocessing 4.1 Know your data 4.2 Missing data 4.3 Outlier detection 4.4 Feature engineering 4.4.1 Feature selection 4.4.2 Scaling the predictors 4.5 Labelling 4.5.1 Simple labels 4.5.2 Categorical labels 4.5.3 The triple barrier method 4.5.4 Filtering the sample 4.5.5 Return horizons 4.6 Handling persistence 4.7 Extensions 4.7.1 Transforming features 4.7.2 Macroeconomic variables 4.7.3 Active learning 4.8 Additional code and results 4.8.1 Impact of rescaling: graphical representation 4.8.2 Impact of rescaling: toy example 4.9 Coding exercises II Common supervised algorithms 5 Penalized regressions and sparse hedging for minimum variance portfolios 5.1 Penalized regressions 5.1.1 Simple regressions 5.1.2 Forms of penalizations 5.1.3 Illustrations 5.2 Sparse hedging for minimum variance portfolios 5.2.1 Presentation and derivations 5.2.2 Example 5.3 Predictive regressions 5.3.1 Literature review and principle 5.3.2 Code and results 5.4 Coding exercise 6 Tree-based methods 6.1 Simple trees 6.1.1 Principle 6.1.2 Further details on classification 6.1.3 Pruning criteria 6.1.4 Code and interpretation 6.2 Random forests 6.2.1 Principle 6.2.2 Code and results 6.3 Boosted trees: Adaboost 6.3.1 Methodology 6.3.2 Illustration 6.4 Boosted trees: extreme gradient boosting 6.4.1 Managing loss 6.4.2 Penalization 6.4.3 Aggregation 6.4.4 Tree structure 6.4.5 Extensions 6.4.6 Code and results 6.4.7 Instance weighting 6.5 Discussion 6.6 Coding exercises 7 Neural networks 7.1 The original perceptron 7.2 Multilayer perceptron 7.2.1 Introduction and notations 7.2.2 Universal approximation 7.2.3 Learning via back-propagation 7.2.4 Further details on classification 7.3 How deep we should go and other practical issues 7.3.1 Architectural choices 7.3.2 Frequency of weight updates and learning duration 7.3.3 Penalizations and dropout 7.4 Code samples and comments for vanilla MLP 7.4.1 Regression example 7.4.2 Classification example 7.4.3 Custom losses 7.5 Recurrent networks 7.5.1 Presentation 7.5.2 Code and results 7.6 Other common architectures 7.6.1 Generative adversarial networks 7.6.2 Autoencoders 7.6.3 A word on convolutional networks 7.6.4 Advanced architectures 7.7 Coding exercise 8 Support vector machines 8.1 SVM for classification 8.2 SVM for regression 8.3 Practice 8.4 Coding exercises 9 Bayesian methods 9.1 The Bayesian framework 9.2 Bayesian sampling 9.2.1 Gibbs sampling 9.2.2 Metropolis-Hastings sampling 9.3 Bayesian linear regression 9.4 Naïve Bayes classifier 9.5 Bayesian additive trees 9.5.1 General formulation 9.5.2 Priors 9.5.3 Sampling and predictions 9.5.4 Code III From predictions to portfolios 10 Validating and tuning 10.1 Learning metrics 10.1.1 Regression analysis 10.1.2 Classification analysis 10.2 Validation 10.2.1 The variance-bias tradeoff: theory 10.2.2 The variance-bias tradeoff: illustration 10.2.3 The risk of overfitting: principle 10.2.4 The risk of overfitting: some solutions 10.3 The search for good hyperparameters 10.3.1 Methods 10.3.2 Example: grid search 10.3.3 Example: Bayesian optimization 10.4 Short discussion on validation in backtests 11 Ensemble models 11.1 Linear ensembles 11.1.1 Principles 11.1.2 Example 11.2 Stacked ensembles 11.2.1 Two-stage training 11.2.2 Code and results 11.3 Extensions 11.3.1 Exogenous variables 11.3.2 Shrinking inter-model correlations 11.4 Exercise 12 Portfolio backtesting 12.1 Setting the protocol 12.2 Turning signals into portfolio weights 12.3 Performance metrics 12.3.1 Discussion 12.3.2 Pure performance and risk indicators 12.3.3 Factor-based evaluation 12.3.4 Risk-adjusted measures 12.3.5 Transaction costs and turnover 12.4 Common errors and issues 12.4.1 Forward looking data 12.4.2 Backtest overfitting 12.4.3 Simple safeguards 12.5 Implication of non-stationarity: forecasting is hard 12.5.1 General comments 12.5.2 The no free lunch theorem 12.6 First example: a complete backtest 12.7 Second example: backtest overfitting 12.8 Coding exercises IV Further important topics 13 Interpretability 13.1 Global interpretations 13.1.1 Simple models as surrogates 13.1.2 Variable importance (tree-based) 13.1.3 Variable importance (agnostic) 13.1.4 Partial dependence plot 13.2 Local interpretations 13.2.1 LIME 13.2.2 Shapley values 13.2.3 Breakdown 14 Two key concepts: causality and non-stationarity 14.1 Causality 14.1.1 Granger causality 14.1.2 Causal additive models 14.1.3 Structural time series models 14.2 Dealing with changing environments 14.2.1 Non-stationarity: yet another illustration 14.2.2 Online learning 14.2.3 Homogeneous transfer learning 15 Unsupervised learning 15.1 The problem with correlated predictors 15.2 Principal component analysis and autoencoders 15.2.1 A bit of algebra 15.2.2 PCA 15.2.3 Autoencoders 15.2.4 Application 15.3 Clustering via k-means 15.4 Nearest neighbors 15.5 Coding exercise 16 Reinforcement learning 16.1 Theoretical layout 16.1.1 General framework 16.1.2 Q-learning 16.1.3 SARSA 16.2 The curse of dimensionality 16.3 Policy gradient 16.3.1 Principle 16.3.2 Extensions 16.4 Simple examples 16.4.1 Q-learning with simulations 16.4.2 Q-learning with market data 16.5 Concluding remarks 16.6 Exercises V Appendix 17 Data description 18 Solutions to exercises 18.1 Chapter 3 18.2 Chapter 4 18.3 Chapter 5 18.4 Chapter 6 18.5 Chapter 7: the autoencoder model and universal approximation 18.6 Chapter 8 18.7 Chapter 11: ensemble neural network 18.8 Chapter 12 18.8.1 EW portfolios 18.8.2 Advanced weighting function 18.9 Chapter 15 18.10 Chapter 16 Bibliography Index