This thesis investigates the intersection of machine learning and big data with financialforecasting and asset allocation. It focuses on the development of model-based approaches for return prediction and portfolio construction, emphasizing adaptive, dataintensive methods that are responsive to market complexity and rich data environment.The first chapter reviews the evolution of quantitative asset allocation, from Markowitz’sModern Portfolio Theory to advanced machine learning approaches. It explains howclassical models face limitations in return estimation and highlights how machine learning methods address these challenges by capturing complex patterns and nonlinearitieswithout rigid parametric assumptions. It also discusses innovations in covariance matrix estimation, from basic sample statistics to sophisticated big data approaches. Thefinal section outlines the research contributions of this thesis across three chapters.The second chapter is based on the paper “Single-stage Portfolio Optimization withAutomated Machine Learning for M6” published in the International Journal of Forecasting. This chapter examines the M6 forecasting competition as a test of the efficientmarket hypothesis. We challenge the conventional “estimate-then-optimize” approachwith one that directly optimizes portfolio weights from data. We formulate asset allocation as a constrained penalized regression problem with automated model selectionand hyperparameter tuning, and show how weights can be optimized using the Methodof Moving Asymptotes. Testing on the M6 dataset, we achieve a 9.5% annual returnand an information ratio of 5.045, far exceeding the M6 benchmark of 0.5% and 0.453.The third chapter introduces a knowledge-based Black-Litterman (KBL) frameworkthat generates experts’ views through a distributed modeling approach. We develop aggregation rules combining the uncertain and incomplete knowledge from experts whilecalibrating their reliability. To improve covariance estimation, we construct realizedcovariances from high-frequency returns using the Conditional Threshold Autoregres3sive Wishart model. We propose an analytical framework that addresses estimationchallenges. Empirically, KBL consistently achieves strong risk-adjusted performance.The improvement over 1/N is statistically significant post transaction costs.The fourth chapter This chapter is based on the paper “The Diversification Benefitsof Cryptocurrency Asset Categories and Estimation Risk: Pre and Post Covid-19”published in European Journal of Finance, 29(7), 2023. This chapter presents the firststudy examining the diversification benefits of cryptocurrency asset categories. Usingmachine learning for parameter estimation, we find most categories offer significant outof-sample diversification benefits, especially for risk-averse investors. Smart Contracts,PoW coins, PoS coins, and Masternode are the strongest diversifiers that consistentlyoutperform the benchmark. The benefits persist across both stable (pre-COVID) anduncertain (post-COVID) economic conditions.The fifth chapter summarizes the contributions of the thesis and outlines four promisingfuture research directions: applying reinforcement learning for dynamic portfolio optimization, developing explainable machine learning models to improve transparency inalgorithmic decision-making, extending the methodologies to global markets and alternative asset classes, and creating multi-objective optimization frameworks that betterreflect real-world investor constraints. The chapter concludes by emphasizing that theprimary source of value lies in the careful integration of computational methods withinfinancially coherent frameworks, creating investment strategies that maintain interpretability while adapting to complex market environments through the combinationof statistical learning, optimization techniques, and large-scale data analysis.
Quantitative Trading with Machine Learning: Forecasting and Portfolio Optimization
Huang, X. (Author). 12 Nov 2025
Student thesis: Doctoral Thesis › PhD