| name | stock-pool-experiment-runner |
| description | Use when orchestrating a J-stock custom stock-pool experiment where a fixed entry condition is applied to a broad/full universe during a training window, good tickers are selected into one or more test universes, and those universes are evaluated out-of-sample against current and full-pool controls. |
Stock Pool Experiment Runner
Coordinate the stock-pool selection workflow without collapsing the train/test boundary.
Atomic Tasks
Route work to these skills in order:
stock-pool-split-planner: define train windows, test windows, controls, pool sizes, and leakage rules.
stock-pool-entry-signal-scan: generate training-period entry-signal-analysis datasets from the broad/full universe.
stock-pool-ticker-selector: aggregate training signals by ticker and export selected universe JSON files.
stock-pool-oos-evaluation-plan: generate an approval-ready OOS evaluation manifest comparing selected, current, and full universes.
evaluation-batch-executor: execute only after the exact manifest is approved.
stock-pool-oos-review: summarize completed OOS results and leakage checks.
Use existing evaluation-batch-runner only when the user asks for a full batch workflow and a manifest approval boundary is needed.
Hard Rules
- Treat ticker selection as model training. Never select tickers using dates that overlap the test window.
- Keep entry condition, ranking strategy, entry filters, sizing, capacity, and exit strategy fixed across universe controls unless the experiment explicitly varies one of them.
- Compare at least: current production pool, broad/full pool, and trained selected pool of the same target size as the current pool.
- Do not edit production config to run experiments. Pass universe files explicitly through CLI args or manifests.
- Use
G:/My Drive/AI-Stock-Sync/config.json as production source of truth when resolving production defaults.
- Use absolute date ranges in plans and reports.
Reporting
Before any long run, report:
- Train window and test window for each fold.
- Universe files for current, full, and selected pools.
- Pool sizes to export, usually including
140 plus sensitivity sizes such as 80 and 200.
- The exact command or manifest that will run next.
- The approval state when execution would start.