Essays in High Frequency Trading, Portfolio Selection and Oil Futures Markets

Abstract

High frequency trading (HFT) requires a detailed analysis of the quote structure of the continuous limit order book in order to correctly de- rive viable arbitrage strategies. Traders can manipulate order books by submitting and retracting ‘spoof’ orders at various levels of the order book by introducing, quote volume at or above (below) the best ask(bid). However, the limit order book data for heavily traded finan- cial instruments presents an almost unique problem to the econome- trician interested in constructing high frequency measures of liquidity impact over and above the inside spread. A single month of data for an individual maturity of an activity traded futures contract, in our example light crude, can easily exceed 10 Billion bytes of data, even when stored using the single precision floating point format.
In this thesis we conduct a large scale analysis of the West Texas Inter- mediate (WTI) futures contract across the 120 simultaneously traded maturities for five levels of the order book from 2008 to 2016 sample at the continuous limit. Using this very-large data-set we estimate a new form of realized vector autoregression and derive the impulse re- sponse functions useful in building a HFT strategy. we show that for WTI futures a speed of execution of the order of 100s of milliseconds is needed to fully exploit a false quoting strategy designed to system- atically unbalance the order flow. Furthermore, we demonstrate that viable strategies can be built by spoofing up to three levels above the inside spread.
A second part of the thesis involves creating new bootstrap routines to extract meaningful composition data to generate factor pricing mod- els from high frequency data. The key element of this analysis is in understanding the eigendecomposition and subsequent principal com- ponent analysis to extract factors from the data. our bootstrap is new and we provide an analysis of power and consistency in correct- ing bias in the estimation of the eigenstructure and hence evaluating the optimal number of principal components within the data.