Reinforcement Learning in Algorithmic Trading

What I investigated

This thesis explored whether reinforcement-learning agents can develop trading behaviour that remains robust across different market periods rather than only producing promising results during training.

What I built

• Several reinforcement-learning-based trading prototypes
• Time-separated training, validation and test environments
• Walk-forward and repeated runs
• Rule-based and constant baseline strategies
• Synthetic learning tasks for controlled experiments

What I learned

Positive backtests alone are not evidence of a robust trading strategy. The experiments highlighted problems around generalisation, reward design, exploration and the transfer of learned behaviour to unseen market periods.

Public Management Summary

The official public summary of the thesis is available through FHNW.