hajirazin/learnfinance-models-sac

LearnFinance SAC Model - v2026-10-02_4e5831c9

SAC v3 masked-attention portfolio allocator with ranked stock tokens.

Model Details

• Version: v2026-10-02_4e5831c9
• Model Type: SAC v3 (masked attention, 30 stock slots + CASH)
• Training Window: 2016-01-01 to 2026-10-02
• Symbols: 15 stocks

Components

• actor.pt - Gaussian policy network
• critic.pt - Twin Q-value networks
• critic_target.pt - Target Q-value networks
• log_alpha.pt - Entropy temperature coefficient
• scaler.pkl - PortfolioScaler for state normalization
• symbol_order.json - Ordered list of portfolio symbols
• metadata.json - SAC v3 schema, architecture, slot map, and audit metadata
• sac_v3_auxiliary.json - HMM parameters/scaler/labels, causal cutoff state, and market tail

Metrics

• Actor Loss: 0.7087951302528381
• Critic Loss: 0.3011290431022644
• Avg Episode Return: 0.2680437052841098
• Avg Episode Sharpe: 0.16980871650842833
• Eval Sharpe: 1.744801662301668
• Eval CAGR: 0.5650465767056425
• Eval Max Drawdown: 0.18415490988733238

Usage

from brain_api.storage.sac import SACHuggingFaceModelStorage
from brain_api.storage.sac.local import SACHalalFilteredModelStorage

storage = SACHuggingFaceModelStorage(
    repo_id="hajirazin/learnfinance-models-sac",
    local_cache=SACHalalFilteredModelStorage(),
)
artifacts = storage.download_model(version="v2026-10-02_4e5831c9")