| name | stock-pool-split-planner |
| description | Use when designing leakage-safe train/test windows, universe controls, and pool-size variants for J-stock experiments that select tickers from historical entry-signal performance before out-of-sample evaluation. |
Stock Pool Split Planner
Create the experiment contract before generating data or running backtests.
Inputs
Resolve or ask for:
- Fixed entry strategy and whether production ranking/filter defaults should be pinned.
- Broad/full universe file to scan during training.
- Current production pool file from
G:/My Drive/AI-Stock-Sync/config.json.
- Target pool sizes, with
140 as the primary size when matching the current pool.
- Train/test folds, using only dates before each test fold to select tickers.
- Whether the selection is entry-only or entry-plus-exit. For the user's current workflow, default to entry-only.
Default Split Shape
Prefer anchored walk-forward style when the user has no stronger preference:
- Train
2022-01-01 through 2024-12-31, test 2025-01-01 through 2025-12-31.
- Train
2022-01-01 through 2025-12-31, test from 2026-01-01 through the latest available date.
If data availability differs, adjust with explicit absolute dates and state the reason.
Controls
Plan these universe controls for each OOS fold:
current140: production/current 140-stock pool, passed as an explicit universe file.full: broad/full universe, with survivorship-bias caveats if it is based on a current listing file.selected_top140: tickers selected from the training window only.- Optional sensitivity pools:
selected_top80, selected_top200.
Output
Return a compact plan table with:
- Fold id.
- Training range.
- Test range.
- Scan universe.
- Output selected-pool files to be generated.
- OOS universe controls.
- Any unresolved inputs.
Do not run entry-signal-analysis, write universe files, or generate evaluation manifests in this skill.
