Rethinking Reinforcement Learning in Finance

In his recent research talk, Mathis Jander presented critical insights from his ongoing research on the use of reinforcement learning (RL) in finance. After reviewing 166 academic publications, he questioned the field's reliance on benchmark testing to validate new RL models.

He explained that current methods often fail to show whether models can generalize across different time periods or financial assets. Researchers typically assume that financial markets have patterns and that RL agents can exploit them. However, these assumptions remain unproven.

Mathis Jander highlighted that positive results in published studies could simply be due to chance rather than true learning. To address this, he suggested a shift away from purely empirical testing toward building a stronger theoretical understanding of when and why RL should work in financial markets.

He concluded that current research practices need to evolve, calling for new methods that can provide stronger, more reliable evidence.

Presentation and Event Information