Automated Trading System

A FinTech startup sought to develop an automated system to manage complex, high-frequency trading strategies across multiple financial instruments.

Extremely competitive market segment. Large hedge-funds with virtually unlimited resources were in the game.
Most of the future and past events did not correlate. Financial markets appeared to be highly stochastic and non-stationary environments.
Exchange fees and bot messages delivery time as well as exchange processing delays had to be considered.
Handle sudden trend changes due to unexpected global events like epidemics, earthquakes, and wars.
Hide deployed agents' activity from bot detection software.
Automate selection of the best trading agent or ensemble the best trading agents.

14 instruments have been integrated into the trading system.
The need for human labor has been drastically reduced: Our team automated feature selection as well as training, testing, selection, and deployment of an agent with optimal trading strategy per instrument.
Automatically deployed agents demonstrated monthly cumulative mean net profit from 18% to 61% in simulation and from 3% to 28% in real trading with stop-loss and take-profit thresholds.

We briefly tested and quickly abandoned classical Supervised Learning approaches like prediction of close price or direction of a price movement (up, down, flat). They did not work as markets were already filled with such bots. We decided to go with technically complex but non-trivial Deep Reinforcement Learning approach.
Market stochasticity and absence of autocorrelation between events have been overcome with custom reward functions. The training set was constructed by sampling periods of high-price activity to regularise the dynamics of the reward functions.
For early testing and validating AI traders' performance, we developed several exchange simulators. The simulators emulated market and limit orders as well as unpreceded global trend changes, plus considered message delays and market fees.
Feature importance and dimensionality reduction techniques have been applied to limit the observation space and select only significant features.
Multiple feature sets and train-test horizons as well as machine learning algorithms with multiple reward objectives are automatically evaluated using historical data. The best candidates are chosen and deployed as active trading agents.