2020/11/29 by Aaron Wray, Wray, Aaron, Matthew Meades +3
Decision Sciences · Economics, Econometrics and Finance · #FOS: Economics and business #Financial Markets and Investment Strategies #Market Dynamics and Volatility #Stock Market Forecasting Methods #Trading and Market Microstructure (q-fin.TR)
paper · pdf · doi:10.48550/arxiv.2012.00821
openalex publication_date 2020/11/29 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
We present results demonstrating that an appropriately configured deep\nlearning neural network (DLNN) can automatically learn to be a high-performing\nalgorithmic trading system, operating purely from training-data inputs\ngenerated by passive observation of an existing successful trader T. That is,\nwe can point our black-box DLNN system at trader T and successfully have it\nlearn from T's trading activity, such that it trades at least as well as T. Our\nsystem, called DeepTrader, takes inputs derived from Level-2 market data, i.e.\nthe market's Limit Order Book (LOB) or Ladder for a tradeable asset. Unusually,\nDeepTrader makes no explicit prediction of future prices. Instead, we train it\npurely on input-output pairs where in each pair the input is a snapshot S of\nLevel-2 LOB data taken at the time when T issued a quote Q (i.e. a bid or an\nask order) to the market; and DeepTrader's desired output is to produce Q when\nit is shown S. That is, we train our DLNN by showing it the LOB data S that T\nsaw at the time when T issued quote Q, and in doing so our system comes to\nbehave like T, acting as an algorithmic trader issuing specific quotes in\nresponse to specific LOB conditions. We train DeepTrader on large numbers of\nthese S/Q snapshot/quote pairs, and then test it in a variety of market\nscenarios, evaluating it against other algorithmic trading systems in the\npublic-domain literature, including two that have repeatedly been shown to\noutperform human traders. Our results demonstrate that DeepTrader learns to\nmatch or outperform such existing algorithmic trading systems. We analyse the\nsuccessful DeepTrader network to identify what features it is relying on, and\nwhich features can be ignored. We propose that our methods can in principle\ncreate an explainable copy of an arbitrary trader T via "black-box" deep\nlearning methods.\n