This dissertation makes several contributions to the literature on Financial Technology. In the first two chapters I push the frontier of Artificial Intelligence (AI) applications in finance by developing a novel Large Language Topic Model and a constrained optimization framework marrying Reinforcement Learning with Natural Language Processing. I demonstrate the value of these algorithms by applying them to mergers and acquisitions and faith-based investing, respectively. In the final chapter I investigate how perpetual contracts, originally proposed by Robert Shiller in 1992 to price illiquid assets, affect microstructure in underlying spot markets using cryptocurrency markets as a sandbox. More specifically, in Chapter 1 I propose a novel Large Language Model to generate time-varying exposure of firms to arbitrary user-defined risks using conference calls. This approach respects language complexity, does not require an expert to generate an extensive list of keywords, produces exposure measures with respect to any query and is directly applicable to any textual input. The model learns from synthetic corpora for the first time in finance research. Leveraging this methodology, I generate firm-level exposures to geopolitical risk, population aging, climate change and demonstrate how exposure to such risks affects M&A activity. Geopolitical risk makes firms more likely to be an acquirer and less a target. It is propagated through supply chains and makes firms acquire targets in the U.S. with higher investment irreversibility driving vertical integration within the U.S. Population aging increases aggregate M&A activity doubling the effect from labor shortage on M&A activity. Finally, I show that managers pay larger premia for low climate change risk targets, while investors react negatively to deal announcements of targets undergoing structural change. Chapter 2 (joint work with Maureen O'Hara) introduces a new approach for Socially Responsible Investing based on techniques from Artificial Intelligence to enhance investor returns. Focusing on faith-based investing, our approach draws on Large Language Models and Deep Reinforcement Learning (DRL) to address the challenges posed by moral constraints on portfolio selection. Using the Global X S&P500 CATH ETF (the largest Catholic values fund) as an example, we use textual analysis to identify companies consistent with the values mandate, allowing us to create “synthetic CATH” portfolios of different sizes and with longer time horizons. We further optimize each portfolio using DRL to arrive at an optimal set of portfolio weights that maximize out-of-sample Sharpe ratios. Using the tools of AI, we demonstrate dramatic improvements in risk-adjusted returns closing the performance gap in values-based investing. Finally, Chapter 3 (joint work with Qihong Ruan) examines perpetual futures contracts' impact on cryptocurrency spot market quality. Using high-frequency order book data from 2017 to 2023, we document that spot market quality follows a U-shaped pattern over perpetual contracts' eight-hour funding cycles. Exploiting both the exogenous termination of perpetual trading at Huobi Exchange and 95 staggered contract introductions, we identify a seemingly puzzling liquidity pattern: perpetual contracts increase spot trading volume while widening quoted spreads. To resolve this puzzle, we demonstrate that this pattern reflects increased informed trading, particularly during funding settlement hours and periods of larger funding fee magnitudes. Market makers respond to heightened adverse selection risk by widening quoted spreads.