Machine Learning for Asset Management

Emmanuel JURCZENKO is Director of Graduate Studies and Professor of Finance at Glion Institute of Higher Education, Switzerland. Prior to this, he spent 13 years as Associate Professor of Finance at ESCP-Europe and worked for ABN-AMRO as Head of Quantitative Analysts where he was in charge of quantitative fund selection. His research focuses on portfolio construction in particular on risk budgeting, factor investing and machine learning estimation techniques.

• Foreword xiiiAcknowledgments xvChapter 1. Time-series and Cross-sectional Stock Return Forecasting: New Machine Learning Methods 1David E. RAPACH and Guofu ZHOU1.1. Introduction 11.2. Time-series return forecasts 31.2.1. Predictive regression 31.2.2. Forecast combination 51.2.3. Elastic net 61.2.4. Combination elastic net 81.3. Empirical application 101.3.1. Data 101.3.2. Forecasts 121.3.3. Statistical gains 171.3.4. Economic gains 231.4. Cross-sectional return forecasts 261.5. Conclusion 291.6. Acknowledgements 301.7. References 30Chapter 2. In Search of Return Predictability: Application of Machine Learning Algorithms in Tactical Allocation 35Kris BOUDT, Muzafer CELA and Majeed SIMAAN2.1. Introduction 352.2. Empirical investigation 382.2.1. The data 382.2.2. Tactical asset allocation strategy 402.2.3. Implementation 412.2.4. Benchmarks 422.3. A review of machine learning algorithms for prediction of market direction 422.3.1. K-nearest neighbors 432.3.2. Generalized linear model 442.3.3. Elastic net regression 442.3.4. Linear discriminant analysis 452.3.5. Support vector machines with radial kernel 452.3.6. C5.0 472.3.7. Random forests 482.3.8. Multilayer perceptron 482.3.9. Model averaging 492.3.10. Repeated k-fold cross validation 502.4. Evaluation criteria 512.4.1. Statistical performance 512.4.2. Financial performance 532.4.3. Significant features 542.5. Results and findings 542.5.1. Descriptive statistics of the data 552.5.2. Statistical performance 562.5.3. Financial performance 582.5.4. The best performer, benchmark and model average 672.5.5. LIME 682.6. Conclusion 702.7. Acknowledgments 702.8. References 70Chapter 3. Sparse Predictive Regressions: Statistical Performance and Economic Significance 75Daniele BIANCHI and Andrea TAMONI3.1. Introduction 753.2. Related literature 783.3. Data: portfolios and predictors 803.4. Econometric framework 843.4.1. Shrinkage priors 863.4.2. Forecast evaluations 923.5. Predicting asset returns: empirical results 933.5.1. Statistical performance 933.5.2. Economic significance 963.6. Discussion on the dynamics of sparsity 1003.7. Conclusion 1023.8. Appendix 1033.9. Posterior simulation 1033.9.1. Ridge regression 1033.9.2. Lasso and group-lasso 1033.9.3. Elastic net 1053.9.4. Horseshoe and the group horseshoe 1053.10. References 106Chapter 4. The Artificial Intelligence Approach to Picking Stocks 115Riccardo BORGHI and Giuliano DE ROSSI4.1. Introduction 1154.2. Literature review 1204.3. Data 1234.3.1. Equity factors 1234.3.2. Data cleaning 1254.3.3. Features used for training and prediction 1254.4. Model specification and calibration 1264.4.1. Models 1264.4.2. Model calibration 1334.5. Predicting US stock returns 1354.5.1. Information coefficients 1364.5.2. Long–short strategy 1384.5.3. Returns correlation with Alpha model 1404.5.4. Active returns by basket 1414.5.5. Calibrated hyperparameters and model complexity 1424.5.6. Variable importance 1444.6. Predicting European stock returns 1464.6.1. Information coefficients 1464.6.2. Long–short strategy 1474.6.3. Returns correlation with Alpha model 1504.6.4. Active returns by basket 1504.6.5. Calibrated hyperparameters and model complexity 1514.6.6. Variable importance 1524.7. The impact of transaction costs 1544.7.1. Optimized strategies for European stocks 1544.7.2. Optimized strategies for US stocks 1584.8. Conclusion 1614.9. References 163Chapter 5. Enhancing Alpha Signals from Trade Ideas Data Using Supervised Learning 167Georgios V. PAPAIOANNOU and Daniel GIAMOURIDIS5.1. Introduction 1675.2. Data 1695.3. Model and empirical design 1745.4. Estimation and robustness 1795.5. Economic significance 1865.6. Conclusion 1885.7. References 189Chapter 6. Natural Language Process and Machine Learning in Global Stock Selection 191Yin LUO6.1. Introduction 1916.1.1. The performance of traditional stock selection factors continues to shrink 1916.1.2. Textual data, natural language processing and machine learning 1956.2. Natural language analysis of company management presentations 1976.2.1. Coverage 1986.2.2. Readability index and language complexity 2016.2.3. Quantifying executive personalities 2066.2.4. Syntactic parser and part-of-speech (POS) tagging 2076.3. Extracting long-term signal from news sentiment data 2116.3.1. Introducing RavenPack data 2116.3.2. The challenges of using news sentiment signals in stock selection 2156.3.3. How do investors react to news? 2166.3.4. The interaction of news, corporate events and investor behavior 2176.3.5. A machine learning approach to extract event-based sentiment 2216.3.6. Welcome to NICE (News with Insightful Categorical Events) 2256.4. References 228Chapter 7. Forecasting Beta Using Machine Learning and Equity Sentiment Variables 231Alexei JOUROVSKI, Vladyslav DUBIKOVSKYY, Pere ADELL, Ravi RAMAKRISHNAN and Robert KOSOWSKI7.1. Introduction 2317.2. Data 2347.2.1. Data construction process 2347.3. Methodology 2407.3.1. Historical beta 2417.3.2. Bloomberg’s adjusted beta 2417.3.3. OLS regression 2417.3.4. Post-LASSO OLS regression 2417.3.5. Random forest model 2427.3.6. XGBoost model 2427.4. Empirical results 2427.4.1. Variable selection 2427.4.2. Forecasting models 2447.4.3. Variable importance 2467.4.4. SHAP values. 2477.4.5. Overall level of feature importance 2487.4.6. Cross-sectional analysis of feature importance 2507.4.7. Time-series analysis of feature importance 2537.5. Constructing market neutral long–short portfolios 2577.6. Concluding remarks 2587.7. References 259Chapter 8. Machine Learning Optimization Algorithms & Portfolio Allocation 261Sarah PERRIN and Thierry RONCALLI8.1. Introduction 2628.2. The quadratic programming world of portfolio optimization 2648.2.1. Quadratic programming 2648.2.2. Mean-variance optimized portfolios 2658.2.3. Issues with QP optimization 2708.3. Machine learning optimization algorithms 2718.3.1. Coordinate descent 2748.3.2. Alternating direction method of multipliers 2798.3.3. Proximal operators 2838.3.4. Dykstra’s algorithm 2898.4. Applications to portfolio optimization 2938.4.1. Minimum variance optimization 2958.4.2. Smart beta portfolios 3018.4.3. Robo-advisory optimization3078.4.4. Tips and tricks 3128.5. Conclusion 3158.6. Acknowledgements 3178.7. Appendix 3178.7.1. Mathematical results 3178.7.2. Data 3238.8. References 324Chapter 9. Hierarchical Risk Parity: Accounting for Tail Dependencies in Multi-asset Multi-factor Allocations 329Harald LOHRE, Carsten ROTHER and Kilian Axel SCHÄFER9.1. Hierarchical risk parity strategies 3329.1.1. The multi-asset multi-factor universe 3339.1.2. The hierarchical multi-asset multi-factor structure 3349.1.3. Hierarchical clustering 3389.1.4. Portfolio allocation based on hierarchical clustering 3429.2. Tail dependency and hierarchical clustering 3439.2.1. Tail dependence coefficients 3449.2.2. Estimating tail dependence coefficients 3459.3. Risk-based allocation strategies 3479.3.1. Classic risk-based allocation techniques 3479.3.2. Diversified risk parity 3489.4. Hierarchical risk parity for multi-asset multi-factor allocations 3529.4.1. Strategy universe 3529.4.2. A statistical horse race of risk-based allocation strategies 3549.5. Conclusion 3609.6. Acknowledgements 3629.7. Appendix 1: Definition of style factors 3629.7.1. Foreign exchange (FX) style factors 3629.7.2. Commodity style factors 3639.7.3. Rates style factors 3649.7.4. Equity style factors 3649.8. Appendix 2: CSR estimator 3659.9. References 367Chapter 10. Portfolio Performance Attribution: A Machine Learning-Based Approach 369Ryan BROWN, Harindra DE SILVA and Patrick D. NEAL10.1. Introduction 36910.2. Methodology 37110.2.1. Matrix algebra representation of selection and allocation effects 37210.2.2. Creating categorical variables from continuous variables 37410.2.3. Optimizing continuous variable breakpoints to maximize systematic attribution 37510.3. Data description 37710.4. Results 37810.5. Conclusion 38510.6. References 386Chapter 11. Modeling Transaction Costs When Trades May Be Crowded: A Bayesian Network Using Partially Observable Orders Imbalance 387Marie BRIÈRE, Charles-Albert LEHALLE, Tamara NEFEDOVA and Amine RABOUN11.1. Introduction 38811.2. Related literature 39111.2.1. Transaction costs and market impact 39111.2.2. Bayesian networks 39211.3. ANcerno database 39411.4. Transaction cost modeling 39611.4.1. Order size 39611.4.2. Order flow imbalance 39811.4.3. Joint effect of order size and order flow imbalance 40011.5. Bayesian network modeling with net order flow imbalance as latent variable 40311.5.1. Bayesian inference 40411.5.2. Bayesian network modeling 40611.5.3. Net order flow imbalance dependencies 40911.5.4. Implementation shortfall dependencies 41311.6. Forecasting implementation shortfall 41511.6.1. Inference of investors’ order flow imbalance given post-trade cost and market conditions 42011.7. Conclusion 42111.8. Appendix A: Garman-Klass volatility definition 42311.9. Appendix B: bid-ask spread and volatility distribution dependencies 42311.10. Appendix C: beta distribution properties 42411.11. Appendix D: net order flow imbalance properties 42511.12. Appendix E: implementation shortfall distribution 42511.13. Appendix F: Hastings-Metropolis algorithm 42611.14. References 427List of Authors 431Index 435Commendations 434