Modeling trade direction

I propose a modeling approach to classifying trades as buys or sells. Modeled classifications consider information strengths, microstructure effects, and classification correlations. I also propose estimators for quotes prevailing at trade time. Comparisons using 2,800 US stocks show modeled classifications are 1-2% more accurate than current methods across dates, sectors, and the spread. For Nasdaq and NYSE stocks, 1% and 1.3% of improvement comes from using information strengths; 0.9% and 0.7% of improvement comes from estimating quotes. I find evidence past studies used unclean data and indications of short-term price predictability. The method may help detect destabilizing order flow.