StyleReplica Alpha Signal Research

Simplified Chinese reference: StyleReplica Alpha 信号研究.

alpha_research.style_replica maintains StyleReplica factors, research classifications, A/B scores, and standard signal artifacts. Final slot allocation, theme quotas, holding buffers, replacement limits, overlap handling, and portfolio weights have moved out of the alpha owner.

Pipeline orchestration and targets.json export belong to strategy_pipeline. The frozen StyleReplica policy belongs to strategy-app. Generic target holdings, backtesting, costs, and capacity belong to portfolio_backtester.

Module entry points

from alpha_research.style_replica import (
    StyleReplicaConfig,
    StyleReplicaSignalGenerator,
    generate_daily_signals,
    map_stock_to_theme,
)

The package no longer exposes StyleReplicaPortfolioConfig, build_style_replica_positions, holding-change builders, or portfolio exposure builders.

Signal layer

StyleReplicaSignalGenerator accepts prices, turnover, market capitalization, industry, and security reference information. It computes two research scores, A and B.

Security eligibility used for backtests or simulations must be dated and include symbol, trade_date, is_st, st_available_from, is_suspended, and list_date. An ST state is known for a decision only when st_available_from is no later than that decision date. Missing availability or a date after the decision leaves eligibility unknown, so the security is excluded from the candidate universe. Legacy daily_clean assets without this field cannot provide the new point-in-time guarantee. The current security name cannot substitute for historical ST status. Without security eligibility input, exploratory scores can still be generated, but eligible_for_backtest and eligible_for_live are false. Historical prices only contribute factors available at the time. The latest date's universe must not be projected backward across the sample.

Score A mainly uses residual volatility, liquidity, market capitalization, 20-day and 120-day momentum, market beta, industry momentum, and optional minute-trading activity. Score B mainly uses volatility convergence, low residual volatility, liquidity, 20-day and 120-day momentum, and an optional Hermite stability measure.

Standard signal artifacts include:

Field	Meaning
`signal_date`	Signal date
`symbol`	Security identifier
`score_a`	Research score A
`score_b`	Research score B
`raw_pred`	Unified research ranking score
`signal_eval`	Evaluation score
`signal_backtest`	Final score consumed by downstream portfolio backtests
`leg`	Candidate leg from research classification
`theme`	Research theme classification, without theme quotas
`industry`	Industry
`model_version`	Alpha model version
`feature_set_id`	Feature-set identifier

The standard signal filename is signals_style_replica.parquet; metadata is stored in signals_style_replica.meta.json.

signal_backtest is a score. It does not represent portfolio returns, executions, or account NAV.

Configuration boundary

StyleReplicaConfig contains alpha research parameters and model/feature identity only. Slot counts, theme quotas, industry limits, buffers, replacement rules, overlap settings, and final weights are outside this configuration.

Strategy parameters are frozen by strategy_app.style_replica.StyleReplicaPolicy and converted into the generic sleeve portfolio specification in portfolio-backtester.

alpha_research.style_replica
    signals
      ↓
strategy_app.style_replica.StyleReplicaPolicy
      ↓
portfolio_backtester.sleeve_portfolio
    positions_by_rebalance

Theme mapping

theme_map maps a security to a research theme. It retains theme keys, display labels, industry mappings, and concept-keyword mappings.

It does not define how many securities to buy in each theme. Theme quotas are strategy policy and belong to strategy-app.

Signal stability

The alpha layer may measure changes in Top-K membership using topk_membership_churn. This metric measures changes in the research signal set.

It does not apply portfolio buffers, weights, industry limits, execution feasibility, or transaction costs. portfolio-backtester calculates formal portfolio turnover after target holdings are generated.

Cross-layer boundaries

Develop new factors, scores, signal artifacts, IC, recency, and other research diagnostics here.

Manage slots, buffers, replacement, overlap, final weights, portfolio costs, and capacity in portfolio-backtester. Manage StyleReplica A/B identity, theme quotas, and frozen strategy versions in strategy-app.

Alpha research code must not import portfolio_backtester or quant_execution_engine at runtime.