Essays on asset pricing, machine learning and econometrics

Theodorou, Konstantinos 2025. Essays on asset pricing, machine learning and econometrics. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF - Accepted Post-Print Version Download (8MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (1MB)

Abstract

Chapter 2: Recent studies on the performance of machine learning prediction of asset prices base their estimation of trading frictions on break-even transaction cost and control for limits-to-arbitrage by excluding large sections of the sample. Arguably, this approach is misguided as it has ramifications which oppose well-established empirical findings. This chapter overcomes both issues by directly modelling trading frictions as a function of the market microstructure and uses optimisation to adjust for the effects of frictions. The evidence suggests that the success of the neural network model of Gu et al. (2020) is not spurious, but it captures unique pricing information, and this chapter identifies under which conditions this can be realised in the market. Chapter 3: Neural network models, despite their striking empirical success in asset pricing, are generally considered ‘black boxes’ which obscure the economic foundations of the relationships they identify. This chapter argues that the ‘black box’ notion is a misnomer and shows how to use the methods established in empirical finance to identify the factors which drive the performance of neural networks. Novel results show that the model extracts significant alpha from episodes of investor overreaction, specifically in two scenarios: (a) shorts stocks with recently improved top line when their bottom line fails to improve; and (b) shorts stocks with recently inflated stock prices due to investor overreaction to positive surprises in financial results when their growth fails to be reflected in their fundamentals. Chapter 4: The validity of the two-pass regression of Fama and Macbeth (1973) depends on the time-dimension, T, being large. When individual assets are used to estimate risk premia the cross-sectional dimension, N, is much larger than T. This chapter explores the levels at which T is small enough to undermine the validity of the two-pass procedure. The investigation employs the asymptotically valid framework of Shanken (1992)–Raponi (2020) and finds that, in univariate regressions the divergence between the two models is statistically significant for time windows up to 10 years in samples containing microcaps and up to 7 years in microcap-free samples. This number is lowered to the 5-year mark in multivariate regressions.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Business (Including Economics)
Date of First Compliant Deposit:	10 July 2025
Last Modified:	10 Jul 2025 10:31
URI:	https://orca.cardiff.ac.uk/id/eprint/179666

Actions (repository staff only)

Edit Item