Abstract
Using a large-scale Deep Learning approach applied to a high-frequencydatabase containing billions of electronic market quotes and transactions forUS equities, we uncover nonparametric evidence for the existence of a universaland stationary price formation mechanism relating the dynamics of supply anddemand for a stock, as revealed through the order book, to subsequentvariations in its market price. We assess the model by testing itsout-of-sample predictions for the direction of price moves given the history ofprice and order flow, across a wide range of stocks and time periods. Theuniversal price formation model is shown to exhibit a remarkably stableout-of-sample prediction accuracy across time, for a wide range of stocks fromdifferent sectors. Interestingly, these results also hold for stocks which arenot part of the training sample, showing that the relations captured by themodel are universal and not asset-specific. The universal model --- trained on data from all stocks --- outperforms, interms of out-of-sample prediction accuracy, asset-specific linear and nonlinearmodels trained on time series of any given stock, showing that the universalnature of price formation weighs in favour of pooling together financial datafrom various stocks, rather than designing asset- or sector-specific models ascommonly done. Standard data normalizations based on volatility, price level oraverage spread, or partitioning the training data into sectors or categoriessuch as large/small tick stocks, do not improve training results. On the otherhand, inclusion of price and order flow history over many past observations isshown to improve forecasting performance, showing evidence of path-dependencein price dynamics.