The LOB Recreation Model: Predicting the Limit Order Book from TAQ History Using an Ordinary Differential Equation Recurrent Neural Network

In an order-driven financial market, the price of a asset is discovered through interaction orders - requests to buy or sell at particular that are posted public limit order book (LOB). Therefore, LOB data extremely valuable for modelling market dynamics. However, not freely accessible, which poses challenge participants and researchers wishing exploit this information. Fortunately, trades quotes (TAQ) arriving top LOB, executing in more readily available. paper, we present recreation model, first attempt from deep learning perspective recreate five levels small-tick stocks using only TAQ data. Volumes sitting predicted by combining outputs from: (1) history compiler uses Gated Recurrent Unit (GRU) module selectively compile prediction relevant quote history; (2) events simulator, Ordinary Differential Equation Neural Network (ODE-RNN) simulate accumulation net arrivals; (3) weighting scheme adaptively combine predictions generated (2). By paradigm transfer learning, core encoder trained on one stock can be fine-tuned enable application other assets same class with much lower demand additional Comprehensive experiments conducted two real world intraday datasets demonstrate proposed model efficiently high accuracy as input.