Evaluation orchestration

Language: English · 简体中文

strategy_pipeline.pipeline.eval coordinates metric calculation, portfolio replay, and result collection for an evaluation window.

It delegates scoring to alpha_research.period_evaluation and position, NAV, turnover, and exposure outputs to portfolio_backtester. This module standardizes inputs, call order, and the result-dictionary shape; it does not define strategy rules.

Integration hooks

from strategy_pipeline.pipeline.eval import (
    _build_period_positions,
    _evaluate_period,
)

These underscore-prefixed functions are internal hooks for the public pipeline's owner integration. Strategy repositories should use the strategy_pipeline.control_plane contracts rather than depending on these internal result fields.

_evaluate_period calculates evaluation metrics first, records portfolio and NAV outputs, then adds scored-data and exposure outputs. Position-building and NAV helpers delegate to portfolio_backtester implementations and pass through the configured allow_live_fallback policy. The function accepts optional permutation-test training and test frames.

Empty evaluation window

When test_df_full is None or empty, _evaluate_period skips evaluation and returns the standard empty result structure. The result preserves numeric metric series, empty frames and date lists, and None for unavailable backtest statistics or positions. The caller can continue building the normal run summary without a separate missing-window schema.

The empty result shape is covered by tests/orchestration/test_pipeline_eval.py.