crunch-numinous 0.1.0

Challenge package — participant-facing model interface and scoring

challenge

Public challenge package.

Primary package: numinous

Participant-facing files

• tracker.py — model interface (participants subclass TrackerBase)
• scoring.py — scoring function for local self-eval
• backtest.py — backtest harness (BacktestClient, BacktestRunner, BacktestResult)
• config.py — baked-in coordinator URL and default feed dimensions
• examples/ — quickstarter model implementations

Backtest usage

from numinous.backtest import BacktestRunner
from my_model import MyTracker

result = BacktestRunner(model=MyTracker()).run(
    start="2026-01-01", end="2026-02-01"
)
result.predictions_df   # DataFrame in notebook
result.metrics           # rolling windows + multi-metric enrichment
result.summary()         # formatted output

# result.metrics includes both rolling windows and portfolio-level metrics:
# {
#   'score_recent': 0.42, 'score_steady': 0.38, 'score_anchor': 0.35,
#   'ic': 0.035, 'ic_sharpe': 1.2, 'hit_rate': 0.58,
#   'mean_return': 0.012, 'max_drawdown': -0.08,
#   'sortino_ratio': 1.5, 'turnover': 0.23,
# }

Data is automatically fetched from the coordinator and cached locally. No coordinator URL or feed configuration needed — baked into the package.

Multi-metric enrichment (IC, hit rate, Sortino, etc.) is computed using the same metrics registry as the coordinator, giving competitors identical feedback locally.

Node-private runtime

• ../node/config/ — CrunchConfig, runtime callables