AgentSkillsCN

trading-feature-pipeline

为本仓库构建并进行第二阶段特征工程输出的合理性检查。当用户希望生成、检查或调试features_core/features_structure的Parquet文件，或无前瞻性的摆动逻辑时使用。

SKILL.md
--- frontmatter
name: trading-feature-pipeline
description: Build and sanity-check phase-2 feature engineering outputs for this repository. Use when the user asks to generate, inspect, or debug features_core/features_structure parquet files or no-lookahead swing logic.

Trading Feature Pipeline

Run feature engineering with repository defaults:

  1. •Activate virtual environment: source .venv/bin/activate.
  2. •Build features: make run-features.
  3. •Verify outputs:
    • •artifacts/processed/features_core.parquet
    • •artifacts/processed/features_structure.parquet

Run quick sanity checks after generation:

  1. •Confirm required core columns: time, atr14, atr_pct, range1, ret1, ret3, ret6, abs_ret1, vol_sma50, vol_ratio
  2. •Confirm required structure columns: time, last_swing_high, last_swing_low, dist_high_atr, dist_low_atr, near_structure
  3. •Confirm non-empty row counts and no critical null bursts in key columns.

Enforce no-lookahead behavior:

  1. •Treat pivot confirmation as delayed by swing_size.
  2. •Use only confirmed swings (pivot_index + swing_size) as available structure state.
  3. •Keep features aligned to bar t only.

When reporting results, include:

  1. •Output row counts and column lists.
  2. •Warmup and swing settings read from configs/config.yaml.
  3. •Any anomalies (NaN spikes, empty output, missing columns) with exact file path.

For concrete check snippets, read references/checks.md.