Using the Epps effect to detect discrete data generating processes

The Epps effect is key phenomenology relating to high frequency correlation dynamics in the financial markets. We argue that it can be used to determine whether trades at a tick-by-tick scale are best represented as samples from a Brownian diffusion, perhaps dressed with jumps; or as samples from truly discrete events represented as connected point processes. This can answer the question of whether correlations are better understood as an emergent time scale dependent property. In other words: Is the Epps effect a bias? To this end, we derive the Epps effect arising from asynchrony and provide a refined method to correct for the effect. The correction is compared against two existing methods correcting for asynchrony. We propose three experiments to discriminate between possible underlying representations: whether the data is best thought to be generated by discrete connected events (as proxied by a D-type Hawkes process), or if they can be approximated to arise from Brownian diffusions, with or without jumps. We then demonstrate how the Hawkes representation easily recovers the phenomenology reported in the literature; phenomenology that cannot be recovered using a Brownian representation, without additional ad-hoc model complexity, even with jumps. The experiments are applied to trade and quote data from the Johannesburg Stock Exchange. We find evidence suggesting that high frequency correlation dynamics are most faithfully recovered when tick-by-tick data is represented as a web of inter-connected discrete events rather than sampled or averaged from underlying continuous Brownian diffusions irrespective of whether or not they are dressed with jumps.