Max Vilarasau Serra

Domain Randomization for Deep Reinforcement Learning Agents for Portfolio Management

This study investigates how Domain Randomization (DR) can make deep reinforcement learning agents more reliable in financial portfolio management. Normally, these agents learn from past market data and risk overfitting to specific situations, which weakens their performance when markets change. By deliberately adding random variations to market simulations—such as different levels of volatility, noise, and price behavior—the agents are forced to adapt to a wider range of conditions. Trained with the Deep Deterministic Policy Gradient (DDPG) algorithm and tested on Dow Jones data, the DR agents achieved higher Sharpe ratios than those trained conventionally. The results suggest that DR works as a kind of “regularizer,” helping AI trading systems avoid overfitting and remain more robust in unpredictable, real-world markets.

Resources on Remote Viewing Research for Original CIA Experiments

This report shares data from a remote viewing experiment, a line of research originally explored by the CIA, where participants attempt to perceive information about distant or unseen targets. While the phenomenon is controversial, it has sparked ongoing debate in psychology and parapsychology. The study examined whether emotional intelligence (EI) relates to better performance in these tasks. Using structural equation modeling, the authors found a positive link between higher EI and more accurate remote viewing “hits.” To promote transparency, they explained details about how EI was measured, why their effect sizes had underestimated variability, and how the unusual statistical results should be interpreted with caution. Overall, the work contributes to clarifying the methods and discussions around anomalous cognition experiments.

Last updated on January 1st 2026