stock-pool-ticker-selector

name	stock-pool-ticker-selector
description	Use when aggregating J-stock training-period entry-signal-analysis or entry-exit-validation CSV artifacts by ticker, scoring ticker robustness, and exporting selected stock-pool universe JSON files for out-of-sample evaluation.

Stock Pool Ticker Selector

Convert training-period signal results into fixed ticker universe files.

Inputs

Use one of these training-only artifacts:

• Preferred entry-only input: entry_signal_analysis_selected_*.csv.
• Alternative entry-only diagnostic input: entry_signal_analysis_candidates_*.csv.
• Entry-plus-exit input: combo_selected_trades.csv from entry-exit-validation.

Require a known training cutoff. Reject or flag rows with signal_date or entry_date after the training end date.

Ticker Metrics

Aggregate by normalized ticker code. Compute at least:

• sample_count
• active_year_count
• win_rate
• avg_return
• median_return
• trimmed_mean_5pct
• p10_return
• p25_return
• top_5pct_contribution_ratio
• mean_without_top_5pct_return

For entry-signal-analysis, use the planned primary return column such as forward_return_3d_pct or forward_return_5d_pct. For entry-exit-validation, use return_pct.

Selection Rules

Default gates:

• Require sample_count >= 20 for short training windows, or >= 30 for multi-year training windows.
• Require signals in at least two distinct years when the training window spans at least two years.
• Rank by robust central performance and downside control, not raw average alone.
• Penalize high top-tail dependence, especially when top_5pct_contribution_ratio > 0.5 or mean_without_top_5pct_return is negative.

Default score order:

1. Higher trimmed_mean_5pct.
2. Higher median_return.
3. Higher p10_return.
4. Higher win_rate.
5. Lower top_5pct_contribution_ratio.
6. Higher sample_count.

Universe Output

Export JSON that src.utils.universe_loader.load_tickers_from_file can read:

{
  "version": "1.0",
  "pool_id": "selected_top140_train_2022_2025",
  "created_at": "YYYY-MM-DDTHH:MM:SS",
  "selection_source": "<training CSV path>",
  "train_start": "YYYY-MM-DD",
  "train_end": "YYYY-MM-DD",
  "target_size": 140,
  "tickers": [
    {
      "code": "7203",
      "rank": 1,
      "score": 1.234,
      "sample_count": 42,
      "active_year_count": 3
    }
  ]
}

Write selected-pool artifacts under G:/My Drive/AI-Stock-Sync/universe/ unless the user requests a repo-local fixture.

Validation

• Reopen each generated JSON with load_tickers_from_file.
• Confirm unique ticker count equals the intended size, or explain why gates produced fewer names.
• Save the ticker score table next to the JSON when practical.
• Do not update production monitor lists or production config.