Doctor of Philosophy
My thesis, which consists of three chapters, aims to either adapt ML to solve important questions in financial markets, or apply state-of-the art ML methodologies that specifically target on economic causal inference problems. The first chapter is adapted from a paper co-authored with Lars Stentoft and published on the Journal of Portfolio Management in 2019. We modified a ML model in order to develop a portfolio tracking strategy that solves a pain point for investors. Many investors try to replicate the performance of certain financial instrument, such as a hedge fund strategy return index for its desirable risk-return profile. However, access to many popular strategies, including hedge fund index replication, is prohibitive due to either high cost or entry barrier. The proliferation of cheap, liquid benchmark-tracking exchange traded funds (ETFs) provides an opportunity to solve such problem. The challenge is to find the ``right'' ETFs, as the number of relevant ETFs in the ETF universe is very large. The conventional ML-assisted dimension reduction technique, which is to constrain the the number of investment instruments of the model is a natural solution. However, a conventional approach it is not ideal in this investment context, where a trade-off needs to be made between the number of investment instruments and statistical loss, i.e. mean squared portfolio tracking error. The standard statistical resampling procedure called cross-validation that optimize the complexity-loss trade-off is not designed to respect model user's economic loss or utility. This chapter proposes an innovative modification to fill in the gap between statistical loss and economic loss, which aggregates various investment costs and tracking deviations. By selecting the replication instruments in a way that is more consistent with his or her economic utility, the investor can save around 60 bps per year, vs the conventional variable selection procedure.
My second and third chapter focus a popular investment strategy in the currency foreign exchange market. Buying currencies with higher interest rates and selling ones with lower interest rates is commonly known as carry trade speculation, where ``carry" refers to the interest rate differential between two countries. Carry trade strategy is the most widely discussed currency strategy in the finance literature, and carry traders are often blamed for causing large swings in currency prices. Since understanding the true mechanism driving currency market is important for both investors and central banks, in the second chapter, I verify this causal relationship using state-of-the art robust causal inference technique, called de-biased machine learning, and cross checking the conclusion with other more conventional casual inference tools. Due to the complexities of financial markets, a simple linear model or a naïve ML model could easily result in model misspecification, which will in turn bias causal parameter estimation and result in misleading conclusions. I found that contrary to what is typically assumed in the literature and popular belief, there is no evidence that carry trade speculation is an important causal factor in driving currency returns. What appears to be carry-trade related currency reaction should be actually attributed to other mechanisms related to the monetary policy change.
Most carry trade related literature has been focusing on rationalizing the carry trade excess returns as a risk premium. Interestingly, what is at odds with the theoretical prediction is the observed puzzle that the carry trade risk premium has disappeared after the financial crisis. In the last chapter, I show that the puzzle could be attributed to an omitted source of currency risk premium. Some currencies are more sensitive to global economic and financial signals, and therefore attract the attention of currency speculators as the global economy transition through different stages of the business cycle. These currencies have higher ``speculation beta", and therefore earn higher speculation risk premium. However, the two risk premia cancel out after the financial crisis. By constructing a conditional carry trade strategy that removes the contamination from the speculation beta, I am able to recover the carry trade risk premium. Constructing the new strategy requires currency ranking based on each currency's speculation beta. In this paper, the speculation beta is estimated using ML to better capture the non-linear relationships between currency returns and a large pool of economic and financial signals, as well as to reduce the dimension of the problem. I also find that high speculation beta countries tend to be the ones that play a more important role in the global economic network.
Summary for Lay Audience
Data-driven Machine Learning (ML) modelling is not as widely accepted in economics as in the broader statistics community, and a large part of it is due to the fact that ML methods do not deliver estimators with formal large sample properties traditionally reported in econometrics papers. However, some newly developed methods at the intersection of ML and econometrics have recently made advances in theoretical results of this type. When applied to causal inference problems in certain context, they could perform better than traditional econometric methods.
Apart from these special cases, ML methods in general could not deliver valid confidence intervals. However we should not dismiss them due to their advantages in in capturing complex data relationships. This is particularly true for problems involving financial markets data.
My thesis, which consists of three chapters, aims to either adapt ML to solve important questions in financial markets, or apply state-of-the art ML methodologies that specifically target on economic causal inference problems.
Wang, Sha, "Adding Data-driven Modelling to Causal Inference And Financial Economics" (2020). Electronic Thesis and Dissertation Repository. 7485.