Bi-Directional Knowledge Transfer For Continual Deep Reinforcement Learning in Financial Trading

The rapid growth in automated financial trading has high-lighted the need for advanced agents capable of exploiting complex patterns in financial markets. From an algorithmic viewpoint, financial trading is a stochastic and dynamic time series problem. Deep Reinforcement Learning (DRL) has shown great promise in addressing this challenge. It naturally aligns with the objective of financial trading-maximizing rewards-without relying on unrealistic theoretical assumptions that do not hold true in such volatile and noisy data. However, the complexity of the problem still presents challenges for conventional DRL algorithms. To overcome these, continual learning agents are crucial for their ability to adjust to changing market conditions. Our approach adapts continual learning techniques for handling dynamic time series, and introduces a novel bi-directional knowledge transfer loss, that balances the trade-off between maintaining knowledge of past patterns, while adapting to new ones.