Alpha-v2 is a TD3 (Twin Delayed DDPG) actor-critic agent with a distributional Q-critic — quantile regression on the return distribution — trained over a multi-symbol crypto trading universe.
The model couples a shared cross-symbol latent encoder with per-task readout adapters (one for actor, one for critic) and a sign-balanced experience replay sampler explicitly designed to prevent direction collapse during long retraining cycles. Live trading is gated behind in-sample → out-of-sample → live promotion.