Date of Degree
Economics | Finance
Forecast, Exchange Rates, Machine Learning, Deep Learning, Asset Pricing
This dissertation consists of three chapters.
In the first chapter I propose a novel approach to forecast risk premia selecting relevant predictors among hundreds of correlated stock characteristics. I adapt a recently developed method from the deep learning literature, Deep Neural Networks with Group Lasso Regular- ization. This method achieves high out of sample R2, and at the same time yields a sparse representation of the characteristics space that allows for interpretability of the otherwise black box deep learning model. For each period, the model chooses a subset of characteris- tics to be relevant for the risk premia forecast. Our method can handle interactions among variables, hence it is superior to other machine learning algorithms typically used for feature selection such as the Lasso. This work adds to the literature that applies Machine Learning to finance, achieving both high accuracy in forecasting returns and adding interpretabil- ity to the otherwise black box model. Many of the previously identified return predictors don’t provide incremental information for expected returns. Nonlinearities are important and interactions between predictors matter.
In the second chapter I study the relationship between commodity prices and nominal bilateral exchange rates. More specifically, I investigate if there is a distinct commodity related driver of exchange rate movements for a number of currencies, including commodity currencies and large economy currencies, and I investigate if there is a distinct exchange- related driver for commodities. This work is part of a growing literature that in the recent years has evaluated and called into question the ability of commodity currencies to forecast commodity prices, and vice-versa. I find strong evidence that when machine learning models are used, a comprehensive list of mostly traded commodities predicts exchange rates for most currencies in my dataset. Moreover, exchange rates predict commodity prices out of sample and beat the random walk benchmark. Results are robust across estimation window sizes.
In the third chapter we study the relationship between economic fundamentals and nomi- nal bilateral exchange rates. We build on previous literature, and evaluate machine learning models to forecast exchange rates using economic fundamentals. Specifically, we follow the methodology and data collection of Li et al. (2014), and find that their results cannot be replicated using updated data. Moreover, we propose using a different approach than Li et al. (2014) to estimate the output gap. Using output-gap estimated based on the Non- accelerating Inflation Rate of Output method improves the out of sample performance of the models compared to the HP-filtering approach typically used in the literature. This implies that estimating output gap using a simple statistical method such as the HP filter is sub-optimal, and better results can be achieved when using a method that is based on economic theory.
Cupallari, Andi, "Applications of Machine Learning and Deep Learning in Macroeconomic and Financial Forecasting" (2020). CUNY Academic Works.
This work is embargoed and will be available for download on Thursday, June 02, 2022
Graduate Center users:
To read this work, log in to your GC ILL account and place a thesis request.
See the GC’s lending policies to learn more.