We carry out a non parametric analysis of financial durations. We make use of an existing algorithm to describe non parametrically the dynamics of the process in terms of its lagged realizations and of a latent variable, its conditional mean. The devices needed to effectively apply the algorithm to our dataset are presented. On simulated data, the non parametric procedure yields better estimates than the ones delivered by an incorrectly specified parametric method. On a real dataset, the non parametric estimator seems to mildly overperform with respect to its parametric counterpart. Moreover the non parametric analysis can convey information on the nature of the data generating process that may not be captured by the parametric specification. In particular, once intraday seasonality is directly used as an explana- tory variable, the non parametric approach provides insights about the time-varying nature of the dynamics in the model that the standard procedures of deseasonaliza- tion may lead to overlook.
