Test of PIN algorithms through simulation

Abstract

I use simulated series of buys and sells to test nine different PIN estimation algorithms implemented in SAS, proc NLP. I conclude that the algorithms do a good job estimating PIN, on average, when applied to series generated from high‐PIN data generating process (PIN>0.2). However, when applied to series generated from zero‐PIN or low‐PIN (PIN<0.2) data generation process, the algorithms yield overstated PIN estimates. Supplying initial values to parameters in the estimation has strong effect on the estimated PIN. Factorization of the likelihood function plays a less important role. Not supplying initial values dramatically overstates PINs for zero‐PIN and low‐PIN data: zero‐PIN data shows PIN as high as 0.26; low‐PIN data tend to have PINs overstated by 0.05 to 0.10. Supplying initial values as in Lin and Ke (2011) eliminates the bias for low‐PIN data but not for zero‐PIN data. Supplying initial values which correspond to the null hypothesis (of no informed trading) yields zero PIN estimates for low‐Pin as well as high‐PIN data.