Targeting market neutrality

Abstract

Neutralizing portfolios from overall market risk is an important part of investment management, particularly for hedge funds. In this paper we show an economically significant improvement in the accuracy of targeting market neutrality for equity portfolios. Key features of the approach are the relatively short forecast horizon of one week and forecasting with realized beta estimators computed using high quality, error corrected, intraday returns. We also find that too long and too short estimation windows result in poor beta forecasts and that the optimal length of estimation window depends on the frequency of return observations.

Keywords:

• Beta forecasting
• Portfolio optimization
• Short-horizon forecasting
• Systematic risk
• Zero-beta portfolios

JEL Classification:

• C53
• G17

Acknowledgements

We thank Henk Berkman, David Feldman, conference participants at the 37th International Symposium on Forecasting in Cairns, Australia and two anonymous referees for helpful comments.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

§ See Sharpe (1964) and Lintner (1965).

¶ https://quantquote.com/

† https://quantquote.com/

‡ See https://quantquote.com/docs/QuantQuote_Minute.pdf for more information.

§ http://www.crsp.com/

† See Todorov and Bollerslev (2010) for an analysis that decomposes realized betas into components, including a jump component.

† A time trend can occur when realized betas are lower in earlier time periods due to the Epps effect (see Epps 1979) which lowers correlations and betas when the return sampling frequency is too high for accurate return measurement.

† For all beta forecasts presented in this paper, we also conducted subsample analyses by splitting the forecast evaluation period into two half-samples, covering 2004–2008 and 2009–2013 separately. The results are qualitatively similar and the results for 10 minute returns are presented in tables 3 and 4.

† The beta forecasting approaches examined above produce similar results when the forecast evaluation loss function used is the MSE. The MSE results are not tabulated due to space constraints but are available upon request.

‡ The median is computed from daily close observations from the index inception, January 1990, to the end of our sample period, December 2013.