Strategy

skills/strategy owns reusable strategy types that convert features, scores, labels, or signals into date × symbol target weights. Concrete public strategy behavior remains under strategies/.

Public API

from skills.strategy import StrategyContext, StrategyResult, WeightGenerator
from skills.strategy.cross_sectional import (
    DynamicFactorWeightConfig,
    ModularBacktester,
    apply_rebalance_schedule,
    combine_factor_scores,
    estimate_factor_weights,
    hold_weights_on_calendar,
    normalize_factor_frames,
    rank_factor_frames,
    top_n_weights,
)
from skills.strategy.time_series import signal_to_single_asset_weights

Boundaries

• contracts.py defines the strategy-neutral target-weight result and generator protocol.
• ports.py defines only the market-data reader needed by current workflows; it does not predeclare persistence or Tracking APIs.
• cross_sectional/ owns reusable ranking, selection, exit, risk-control, and modular research types, including equal-rank, equal-vote, rolling IC, rolling ICIR, and correlation-aware maximum-ICIR factor combinations.
• hold_weights_on_calendar maps signal-day target weights onto the full trading calendar (forward hold, flat before the first signal) before backtesting.
• time_series.py owns signal-to-weight conversion and a research-only TimeSeriesBacktester adapter that delegates execution to VectorBacktester.
• Concrete factors, features, rules, model pipelines, and workflows belong in strategies/.
• Formal public execution always passes target weights to skills.backtest.VectorBacktester.

TimeSeriesBacktester keeps exploratory prediction-frame analysis available, but does not implement a second return or metric engine. Published public strategy results should still use explicit target weights plus VectorBacktester.

Multi-factor combination recipe

from skills.strategy.cross_sectional import (
    DynamicFactorWeightConfig,
    combine_factor_scores,
)

config = DynamicFactorWeightConfig(
    availability_delay=signal_lag + horizon,
    lookback=252,
    min_periods=126,
    max_weight=0.5,
    correlation_shrinkage=0.5,
)
result = combine_factor_scores(
    raw_factors,
    method="max_icir",
    directions=factor_directions,
    normalization="rank",
    top_n=3,
    ic_history=ic_history,
    correlation_history=rolling_rank_correlations,
    dynamic_config=config,
)

Supported methods are equal rank, equal vote, rolling IC, rolling ICIR, and correlation-aware maximum ICIR. result.factor_weights is the factor-level voice in the composite score; result.target_weights is the separate asset allocation produced after Top-N selection. Maximum ICIR requires tidy correlation history with columns eob, factor_a, factor_b, and correlation. The public combination entry point always applies factor direction and daily cross-sectional normalization. Use normalization="rank" for robust percentile ranks or normalization="zscore" to retain relative score distance; do not pre-normalize inputs.

Time-series recipe

from skills.backtest import VectorBacktester
from skills.strategy.time_series import signal_to_single_asset_weights

weights = signal_to_single_asset_weights(signal, symbol="SHSE.510300")
result = VectorBacktester(
    panel,
    signal_lag=1,
    commission=0.0002,
    slippage_bp=2.0,
).run(weights)