Detecting intentional herding: what lies beneath intraday data in the Spanish stock market

Abstract

This paper examines the intentional herd behaviour of market participants, using Li's test to compare the probability distributions of the scaled cross-sectional deviation in returns in the intraday market with the cross-sectional deviation in returns in an ‘artificially created’ market free of intentional herding effects. The analysis is carried out for both the overall market and a sample of the most representative stocks. In addition, a bootstrap procedure is applied in order to gain a deeper understanding of the differences across the distributions under study. The results show that the Spanish market exhibits a significant intraday herding effect that is not detected using other traditional herding measures when familiar and heavily traded stocks are analysed. Furthermore, it is suggested that intentional herding is likely to be better revealed using intraday data, and that the use of a lower frequency data may obscure results revealing imitative behaviour in the market.

Keywords:

• behaviour
• finance
• time series

Acknowledgements

The authors are grateful to the anonymous referee for the helpful comments and suggestions with respect to the first version of our paper.

N Blasco and S Ferreruela wish to acknowledge the financial support of the Spanish Ministry of Education and Science (SEJ2006-14809-C03-03/ECON), the Spanish Ministry of Science and Innovation (ECO2009-12819-C03-02), ERDF funds, the Caja de Ahorros de la Inmaculada (Europe XXI Programme) and the Government of Aragon. P Corredor is grateful for the financial support of the Spanish Ministry of Education and Science (SEJ2006-14809-C03-01), the Spanish Ministry of Science and Innovation (ECO2009-12819-C03-01), ERDF funds and the Government of Navarra.

Notes

1 This proposal has been used to model diverse phenomena in which bits of information, interacting in pairs, produce collective effects. Although this model is usually acknowledged to usefully explain statistical mechanics, Schneidman et al (2006) show that the Ising model is useful for any model of neural function. They find that collective behaviour is described quantitatively by models that capture the observed pairwise correlations and predict that larger networks are completely dominated by correlation effects.

2 As mentioned before, under the assumption of, at most, weak imitation, the return time series of the notional index should behave as a Gaussian distribution. The value of the Kolmogorov–Smirnov (K–S) test (Chakravarti et al, 1967) is 0.0327 with a p-value of 10.88%, indicating that we cannot strongly reject the normality of the return distribution.

3 We have used either the Epanechnikov and the Gaussian kernels. We have also used several bandwidth parameters. We first follow the Silverman option that suggest a data-based automatic relation h₁=0.9kn^−1.5min (α,π/1.34), Then we used lower values up to h₂=h₁/100. The larger the bandwidth, the smoother the estimate. We always choose the lower h value for each couple of density functions. The results do not vary significantly, and are available upon request.

4 For clarity, the table only captures the central range of interval limits, as we consider them the most relevant in our analysis. Further details on additional intervals can be provided by the authors upon request.

5 The analysis has been repeated changing the interval width (both 0.5 and 2 units). The results do not change significantly and are available upon request.