Transfer entropy calculation for short time sequences with application to stock markets

Abstract

We investigate the estimation of transfer entropy (TE) for short time sequences by correlation-dependent balanced estimation of diffusion entropy employed in the transfer entropy (CBEDETE) method and the normal transfer entropy (NTE) method. Our finding shows that the CBEDETE method is more effective than the NTE method on TE calculation for short time series. Based on this conclusion, we use 38 important stock market indices from 4 continents to create successive financial networks with 10∼60-day windows and 1-day step by the CBEDETE method. By extracting the evolution characteristics of out-/in-degree of stock networks, we obtain the most influential stocks RTS, KOSPI, PSI, NIKKE and AORD of Europe, Asia and Oceania and the most influenced stocks IBOVESPA, NYSE, NASD and MERV of America. Finally, by monitoring the ratio of link numbers of each network and smoothing the curves, we find an interesting result that almost all effective peaks in the smoothed ratio curves are prior to the financial crises, such as the global financial crisis in 2008, China’s stock market crash in 2015, etc.