The key thread running through this thesis is predictability and how it relates to asset pricing and portfolio construction.

Chapter 1, co-authored with Oliver Linton, tests for predictability in asset pricing model residuals to check model specification. We estimate three consumption-based asset pricing models and derive ex-ante expected stock market returns from them. For each model, a suite of tests rejects the null that the model residual, the difference between the ex-ante expected market return and the actual return, is a martingale difference sequence. The ability of these models to explain the own-history predictability of the market return is therefore rejected. Further tests show that lagged returns have too much predictive power over current returns to be consistent with the state variables which explain the market return being the same as the state variables which explain the market return in any of the three models.

Chapter 2 focusses on a specific type of predictive information. I examine whether regulator-required public disclosures of large net short positions can be profitably used to build portfolios. These disclosures do not form the basis of a profitable trading strategy for UK stocks. Long-short portfolios based on these disclosures typically make a profit, but it is statistically insignificant. While certain long-only unit initial outlay portfolios can reliably significantly outperform the market, this outperformance is economically modest: about one percentage point a year in gross and risk-adjusted terms.

Finally, Chapter 3 considers how best to use predictive information. Using predictive information unconditionally optimally produces better portfolios than using the predictive information conditionally optimally. Unconditionally optimal portfolios have higher Sharpe ratios and certainty equivalents, plus lower turnover, leverage, losses and drawdowns than conditionally optimal portfolios. Moreover, the unconditionally optimal portfolios tend to stochastically dominate the conditionally optimal portfolios once transaction costs are accounted for. However, whether unconditionally optimal portfolios are preferred to minimum variance or 1/N portfolios depends on the asset universe.