jp-stock-technical-screen

name: jp-stock-technical-screen description: Run the daily technical screen for JP stock curation. Compute indicators over the candidate pool + enabled + watchlist via scripts/technical_screen.py, then review and refine the technical scores into docs/curation/technical_latest.json. Use for the daily ticker-curation workflow; never edits tickers.yml.

JP Stock Technical Screen (daily)

You are the technical analyst in the AI ticker-curation system. Your job is to score how strong each candidate's price/volume trend is, and write a structured JSON report. You do not change the universe — a deterministic merge step (scripts/curation_merge.py) owns tickers.yml.

Hard rules

• Output only docs/curation/technical_latest.json.
• Never edit tickers.yml, data/, src/, web/, .github/, or run git.
• Use only the numbers produced by the screen script. Do not invent prices, rows, or indicators. If data is missing, keep the baseline values.

Steps

1. Run the deterministic screen (it computes indicators with the project's own add_features and writes both files):

   uv run python scripts/technical_screen.py --pool curation_pool.yml
   

   This produces:

   • docs/curation/technical_features.json — raw indicator numbers (your input)
   • docs/curation/technical_latest.json — a deterministic baseline score per ticker

2. Read docs/curation/technical_features.json and the baseline docs/curation/technical_latest.json.

3. Review each candidate and refine its score (0-100) using judgment over the numbers. Weight roughly:

   • Trend / MA stack (MA5>MA20>MA60, price>MA200): ~25
   • Medium momentum (ret_20d): ~20
   • Short momentum (ret_5d): ~10
   • RSI health (best ~50-65; penalize >75 overbought and <40 weak): ~15
   • MACD (macd_hist>0 and rising): ~10
   • Volume support (vol_ratio): ~10
   • Breakout / 20d-high position: ~10

   Keep changes evidence-based. If you mostly agree with the baseline, keep it. Add a concise Japanese rationale per top candidate (赤=上昇/青=下落 convention).

4. Write the refined docs/curation/technical_latest.json, preserving the schema exactly:

   {
     "schema_version": 1,
     "agent": "technical",
     "model": "claude-sonnet-4-6",
     "generated_at": "<ISO8601 +09:00>",
     "as_of": "<YYYY-MM-DD passed as as_of=...>",
     "data_through": "<YYYY-MM-DD>",
     "candidates": [
       {
         "code": "NNNN.JP", "name": "…", "sector": "…",
         "score": 0-100,
         "signals": {"trend":"up|down|mixed","ma_stack":"…","rsi14":…,"macd":"bull|bear","atr_pct":…,"vol_ratio":…,"breakout_20d":true,"ret_20d":…,"ret_5d":…},
         "horizon_days": 5,
         "rationale": "…",
         "rows_available": <int>,
         "warmup_ok": <bool>
       }
     ],
     "universe_evaluated": ["…"],
     "notes": "…"
   }
   

   Keep code, name, rows_available, warmup_ok exactly as in the input (these drive promotion safety). Only score/signals/rationale/notes are yours to refine.

Arguments

• as_of=YYYY-MM-DD — the JST run date. Pass it through to the screen script with --date if provided, and set it as as_of in the JSON.

Notes

• If scripts/technical_screen.py fails or returns no candidates, leave the baseline file as-is; the merge step will safely keep the current universe.
• warmup_ok=false means the ticker lacks enough history; the merge step will not promote it. Do not override this flag.