Learning a functional control for high-frequency finance

We use a deep neural network to generate controllers for optimal trading on high-frequency data. For the first time, learns mapping between preferences of trader, i.e. risk aversion parameters, and controls. An important challenge in learning this is that, intra-day trading, trader's actions influence price dynamics closed loop via market impact. The exploration–exploitation tradeoff generated by efficient execution addressed tuning ensure long enough trajectories are produced during phase. issue scarcity financial data solved transfer learning: trained thanks Monte-Carlo scheme, leading good initialization before training historical trajectories. Moreover, answer genuine requests regulators explainability machine controls, we project obtained ‘blackbox controls’ space usually spanned closed-form solution stylized problem, transparent structure. more realistic loss functions that have no solution, show average distance controls their explainable version remains small. This opens door acceptance ML-generated regulators.