Three essays on machine learning and time series applications on finance: Skew index and return predictability

Drawing on insights from three interconnected chapters, this Ph.D. thesis delivers a thorough examination of financial market forecasting and predictability, with a special emphasis on the predictive power of the Skew Index, the application of both traditional econometrics and state-of-the-art deep learning models for its forecasting, and the utilization of machine learning and deep learning techniques for predicting returns of the SPDR S&P500 ETF Trust. The research undertakes an exploration of financial indexes, fear indexes, commodities, and ETFs as predictive tools, assessing the effectiveness of classical econometrics and modern deep learning models for forecasting. We study return predictability using machine learning and deep learning techniques. This work not only scrutinizes the predictive accuracy or low error metrics of these methodologies but also highlights their significance in enhancing our understanding of financial market behaviors and improving investment strategies. The first chapter, “Skew Index: Descriptive Analysis, Predictive Power, and Short-term Forecast,” establishes a foundational analysis of the Skew Index within the context of financial market indicators. It focuses on the Index’s capacity to evaluate market tail risks and its predictive ability for the S&P500, applying Ordinary Least Squares (OLS), Generalized Least Squares (GLS), and Generalized Method of Moments (GMM) methodologies. This analysis incorporates the Skew Index alongside other indicators, such as the Volatility Index (VIX), BullBear Spread, Intraday Volatility Index (IVX), and Gold Index (XAU). Results from OLS estimations indicate significant predictive power for both the Skew Index and VIX, as evidenced by the statistically significant t-test p-values for their coefficients. However, the GMM results cast doubt on the VIX’s reliability as a predictor. In contrast, the OLS estimation identifies the IVX and XAU, along with the Skew Index and VIX, as effective predictors, highlighting their utility in analyzing financial markets. The findings indicate the Skew Index’s effectiveness as a fear gauge. The short-term forecasting utilizes ARIMA and GARCH models, focusing on the monthly log returns of the Skew Index. We examined various GARCH variants—standard GARCH, EGARCH, GJR-GARCH, and IGARCH—each paired with different innovation distributions, including Normal, t-Student, and Skew t-Student. Notably, the MA(4)-EGARCH(1,3) with t-Student innovation model emerges as a suitable framework for forecasting the Skew Index’s monthly levels. This analysis is essential for grasping the complexities of financial markets, especially for forecasting and reducing the fallout from economic downturns. Building upon the foundational analysis of the Skew Index, the second chapter, “Skew Index: Machine Learning Approach,” transitions from traditional econometric models to the forefront of technological advancement in financial forecasting: deep learning models. This study explores the Skew Index’s daily levels through neural network architectures (Dense, LSTM, GRU, CNN, CNN-LSTM, and CNN-GRU) and two different sets of input data: Stand-alone and models with external variables. A relevant pillar of this research is the relevance of fine-tuning parameters and hyper parameters, such as the number of nodes, CNN-filters and kernels, the initial learning rate, and the proper activation functions, and optimization rules and functions. For further analysis, we employ simulations to forecast the Skew index daily data while not finetuning the hyperparameters but maintaining the most found in the literature, such as the number of epochs to train the model (Kijewskia & Ślepaczuk, 2020; Lim & Lundgren, 2019; Sethia & Raut, 2019) and a non-varying learning rate (Ghosh et al., 2021; Girsang et al., 2020; Zou & Qu, 2020). We demonstrate the superior forecasting capabilities of finetuned LSTM for both stand-alone and models with external variables and the promising potential of Hybrid-CNN models. Even though simulations using commonly accepted hyperparameters showcased lower performance compared to the finetuned methodology, they pave the way for further exploration of probabilistic models in Deep Learning, including Bayesian neural networks. This shift towards deep learning approaches not only implies a methodological evolution but also broadens the scope of financial risk management strategies. In the third chapter, titled “Return Predictability: Deep Learning Approach,” we implement a suite of machine learning techniques—including Logistic Regression, Ridge Classifier, CatBoost Classifier, Decision Tree Classifier, Naïve Bayes, Linear Discriminant Analysis, and AdaBoost Classifier—alongside advanced deep learning architectures such as Dense, LSTM, GRU, CNN, CNN-LSTM, and CNN-GRU, to examine stock return predictability. This study specifically examines the performance of these models on the SPDR S&P500 ETF Trust (SPY) returns, considering various risk aversion thresholds and a broad spectrum of financial indicators, drawing on frameworks established by Boudt et al., (2020), Hull & Qiao (2017), and various fear indexes. Our analysis reveals that Recurrent neural network (RNN) and Convolutional neural network (CNN) models demonstrate capability in capturing complex market dynamics and showcasing superior investment cumulative returns. In particular, the GRU model, at a risk aversion threshold of τ= 0.2 achieved the higher accuracy of all models while presenting uniform metrics. This could be attributed to their memory-based architecture and relative simplicity when compared to other sophisticated models like LSTM and various Hybrid-CNN models. Adopting a trading strategy that involves increasing our portfolio by 1% or trimming it by 0.5% based on the model’s forecasts, we find that all the examined machine learning and deep learning models can outdo the market’s cumulative return. This underlines the significance of employing machine learning and deep learning predictive model’s adept at navigating varying market conditions and catering to diverse investor risk preferences. Moreover, it accentuates the continuing importance of simpler machine learning models — including Logistic Regression, Ridge Classifier, CatBoost Classifier, Decision Tree Classifier, Naïve Bayes, Linear Discriminant Analysis, and AdaBoost Classifier — in the sphere of return prediction, highlighting their sustained pertinence in the analysis of financial markets. Collectively, these chapters illuminate the critical role of innovative analytical tools and methodologies in enhancing the predictability and understanding of financial markets. The integration of traditional financial indexes with cutting-edge machine learning and deep learning techniques offers a holistic view of market behavior, fostering more informed and strategic investment decisions. This Ph.D. thesis not only contributes to academic discourse but also has practical implications for investors, policymakers, and regulators in navigating the intricacies of global financial markets.

• Autores: Vanegas Herrera, Esteban Nicolás
• Tipo de recurso: Doctoral thesis
• Fecha de publicación: 2024
• Institución: Universidad de los Andes
• Repositorio: Séneca: repositorio Uniandes
• Idioma: eng
• OAI Identifier: oai:repositorio.uniandes.edu.co:1992/74576
• Acceso en línea: https://hdl.handle.net/1992/74576
• Palabra clave: Skew Index; SPY; Forecast; Predictability; Machine Learning; Deep Learning; Dense; LSTM; GRU; CNN; Hybrid-CNN; Administración
• License: http://purl.org/coar/access_right/c_f1cf