Files
Abstract
This thesis presents two complementary studies that advance the understanding and application of machine learning techniques in time series forecasting, with a particular focus on financial markets. A comprehensive survey identifies top-performing techniques across tree-based models, deep learning architectures, and hybrid approaches. Building on these insights, the thesis applies a specialized forecasting framework to the domain of day trading. By leveraging a combination of LightGBM models with an extensive set of engineered features ranging from multi-timeframe technical indicators to contextual stock attributes, the model uses two years of second-by-second trade and quote data to estimate risk-reward ratios over multiple forward time horizons. Simulated results using realistic execution constraints demonstrate a pronounced performance advantage over human day traders, yielding daily returns several orders of magnitude higher.