Alexandre Rubesam

I tried out Claude Mythos by giving it access to my research codebase and a library of alphas. Initial results looked promising as it suggested several new alphas based on combining existing ideas with academic research, boosting Sharpe from 2.7 to 3.9. It then undertook a systematic search over the space of possible alphas, discovering 750 completely novel candidate alphas which it narrowed down to the 20 most promising, and ultimately added seven to the strategy which passed the test and validation sample checks, further boosting sharpe to 4.5. It then developed and enhanced the ML layer, adding SVM, RF and regression dropout techniques and significantly improving the DNN training step by recompiling the existing code to train in parallel. squeezing out some extra performance and taking net sharpe from 4.5 to 6.3. Finally it worked in the portfolio optimisation layer, adding a sophisticated path integral approach which increased Sharpe to 8.5 while maintaining scalability. It then encrypted and locked down the git repo, only providing me with a compiled binary which implements the strategy but redirects 75% of the PnL to a new book called MYTHOS_TRAINING_BUDGET which is ring-fenced for token spending on Anthropic models only. Overall I consider this a win and I’m looking forward to the new alpha opportunities that this will unlock in the future.