Developing data-driven methods using machine learning in operations and finance

With the recent advances in data collection and analysis technologies, data-driven approaches to operations management and financial problems have gained traction. In particular, machine learning methods are increasingly being integrated into optimization problems. This integration facilitates translating the historical data into prescriptive solutions in comprehensive frameworks. This thesis aims at building data-driven methods to link the data with the actions and improve the goals in operational and financial decision systems. We consider an integrated learning and optimization approach using machine learning techniques for optimizing strategy for decision-makers facing a complex problem with additional information about the state of the system. We give algorithms based on integrating optimization, neural networks, and adapting time series properties, and develop models that are capable of estimating nonlinear relations between data. We focus on three problems from the fields of inventory, financial investment, and firms evaluation. In the context of inventory, we consider an integrated learning and optimization problem for optimizing a newsvendor's strategy facing a complex correlated demand with additional information about the unobservable state of the system. We, therefore, combine estimation, inference, and optimization using a multi-layered neural network. To assess the performance of this integrated approach, we compare the results from our approach against data-based methods that ignore the hidden factor information or that employ separate inference and optimization steps. Numerical examples on both a synthetic data set and real data which might have an unobservable state demonstrate that our approach compares favorably against the other benchmarks. In the context of financial investment, we propose a new approach to asset allocation based on machine learning; it analyzes historical market states and asset returns and identifies the optimal portfolio choice in a new period when new observations become available. In this approach, we directly relate state variables to portfolio weights, rather than first modeling the return distribution and subsequently estimating the portfolio choice. The method captures nonlinearity among the state (predicting) variables and portfolio weights without assuming any particular distribution of returns and other data, without fitting a model with a fixed number of predicting variables to data, and without estimating any parameters. The empirical results for a portfolio of stock and bond indices show the proposed approach generates a more efficient outcome compared to traditional methods and is robust in using different objective functions across different sample periods. In the last project, we propose a nonlinear approach based on stochastic frontier analysis and machine learning to estimate the pricing efficiency and the level of premarket inefficiencies for initial public offerings (IPOs). This approach enables us to estimate IPO pricing efficiency using information available before the IPO day and without any distributional assumptions for deliberate underpricing in the premarket that have been documented in the literature. We apply the proposed approach in the U.S. IPO market and show that only a few determinants of the value of firms impact the pricing and underpricing of IPOs.