Artificial intelligence, AI, has received increasing attention from the finance industry over recent years. There have been many successful applications of AI in finance, including but not limited to derivative pricing, asset management, credit risk, algorithmic trading, and simulation of time series with stylized facts. This thesis introduces various applications of AI in two major fields, namely (a) quantitative/computational finance and (b) asset management. In each chapter, we address non-stationarity of markets under consideration and focus on building methodologies that would work under the non-stationary behavior of those market.
In computational finance, fast and accurate algorithms are of great importance, especially when analytical solutions are unavailable. Although traditional methods are reliable and easily explainable, they are computationally expensive. Transform methods like Fast-Fourier transform provide faster option pricing, yet they cannot be applied to path-dependent products. In Chapter 2, we build a pricing engine based on supervised deep neural networks. We show that neural networks can replicate major stochastic processes with or without stochastic volatility in both pure diffusion and pure jump frameworks. We validate our models across different ranges of model parameters. Supervised neural networks accelerate the derivative pricing significantly compared to traditional methods.
Applications of AI in asset management are triggered by different dynamics, but they are fully data-driven and thus rely on the availability of data. Chapter 3 proposes a novel prediction framework for cash flow forecasting of illiquid products/assets. Our single-step neural network model provides the investors and managers of funds with a tool to manage the liquidity of their cash flows for financial planning. Our framework is also sensitive to adverse market conditions that could help prepare for upcoming crises such as Covid. In Chapter 4, we propose novel methodologies for mergers and acquisitions (M&A) to predict the deal announcement based on rumors and takeover success. M&A data is highly imbalanced in nature, and the cost of misclassifying a cancelled rumor/cancelled deal as announced deal/takeover success is higher than the other. Hence, we utilize sequential model-based optimization with tree-parzen estimators to maximize the recall score by tuning hyperparameters of neural networks. We improve the recall by 10% without sacrificing accuracy, and our results show that the proposed methodology is robust against changing market environments. In the last chapter, we build a two-step neural network model for sector rotation strategies using macroeconomic variables. The portfolio built based on our proposed model not only beats the benchmark portfolio but can also predict longer horizons.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/hh1a-vw17 |
| Date | January 2023 |
| Creators | Karatas, Tugce |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0019 seconds