Research
HB Capital runs autonomous machine-learning research loops inspired by Andrej Karpathy’s autoresearch ideas, then evaluates candidates against out-of-sample trading evidence.
Continuous search
The system iterates on architecture, reward shaping, and runtime configuration.
Fixed evaluation lane
Candidates are compared on bounded out-of-sample evidence instead of cherry-picked charts.
Frontier tracking
The best-so-far line makes it obvious whether search is climbing or just producing noise.
Research frontier
Public run telemetry is curated before publication.
Framework
The public research surface exposes the live frontier, while the internal app drives launches, lineage, and evaluation evidence.
Candidates are judged from tracked reward logic, training-history metrics, and bounded evaluation evidence. The goal is durable trading behavior, not leaderboard overfitting.