AbstractStatistical arbitrage refers to a suite of quantitative investment strategies employed chiefly by hedge funds and proprietary trading firms. A statistical extension of its pure arbitrage analogue, statistical arbitrage seeks to identify and exploit temporal mis-pricings between two or more securities whose dynamic evolution shares some common stochastic trend. The arbitrageur can draw on a number of different approaches to accomplish this, though the literature is broadly segmented by the distance, cointegration and time series perspectives.
Since the initial academic investigation of statistical arbitrage, its profitability has continued to diminish as the proportion of non-convergent opportunities increased, leading to the hypothesis that spread non-convergence is the cause of declining profitability. This thesis surveys the existing literature, with particular emphasis given to evidence of statistical arbitrage failure, before presenting an approach aimed at unifying the distance, cointegration and time series perspectives under a single explicit model.
The failure of statistical arbitrage opportunities is shown to be the direct consequence of implicit model assumptions that are inconsistent with the empirical literature. An alternative model, the TVHR model, is proposed with the objective of correcting spread non-convergence. A further extension of the model to consider statistical arbitrage profitability in the presence of conventional volatility and unconventional latent regimes is also investigated, offering a comparative analysis of the strengths and weaknesses of each methodology.
This thesis concludes that the declining profitability of statistical arbitrage is not attributable to spread non-convergence, but rather to the distance approach pair selection procedure. The cointegration approach presented in this thesis, by contrast, and the proposed TVHR model variant in particular, halt and even reverse the trend of declining profitability in recent history. This thesis also finds evidence to conclude that statistical arbitrage returns are at least partially dependent on the prevailing volatility regime, and that statistical learning models are better equipped than conventional models to capture and detect latent market regimes.
|Date of Award||13 Oct 2021|
|Supervisor||Bruce Vanstone (Supervisor) & Joachim Hahn (Supervisor)|