A piecewise linear model for trade sign inference

We use transaction level data for twelve stocks with large market capitalization on the Australian Stock Exchange to develop an empirical model for trade sign (trade initiator) inference. The new model is a piecewise linear parameterization of the model proposed recently in Ref. [1]. The space of the predictor variables is partitioned into six regions. Signs of individual trades within the regions are inferred according to simple and interpretable rules. Across the 12 stocks the new model achieves an average out-of-sample classification accuracy of 74.38% (SD=4.25%), which is 2.98% above the corresponding accuracy reported in Ref. [1]. Two of the model's regions, together accounting for 16.79% of the total number of daily trades, have each an average classification accuracy exceeding 91.50%. The results indicate a strong dependence between the predictor variables and the trade sign, and provide evidence for an endogenous component in the order flow. An interpretation of the trade sign classification accuracy within the model's regions offers new insights into a relationship between two regularities observed in the markets with a limit order book, competition for order execution and transaction cost minimization.