# Winsorize

> 对极端值执行缩尾处理并验证影响（outlier 缩尾）

- Skill: `zafer-liu/winsorize` (Agent Skill)
- Install (CLI): `npx skillmds add zafer-liu/winsorize`
- Raw SKILL.md: https://api.skillmd.com/api/skills/zafer-liu/winsorize/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: Zafer-Liu (https://skillmd.com/u/zafer-liu)
- Updated: 2026-09-09
- Page: https://skillmd.com/skills/zafer-liu/winsorize

---

# 缩尾处理

识别目标数值字段和异常分布，说明上下界规则后再缩尾。不得默认覆盖原始数据；比较处理前后的分布、均值和关键指标，并记录边界与受影响行数。

## Tool routing

1. Use `get_schema` to identify the target table and candidate numeric columns.
2. Use `profile_data` to quantify distribution tails before modification.
3. Use `clean_data` with the winsorize operation only after the percentile bounds are clear.
4. Use `query_data` after cleaning to verify clipped values, row counts, and key metric changes.

## Implementation reference

- Tool entries: `agent/tools/business/data.py::_tool_profile_data`, `agent/tools/business/data.py::_tool_clean_data`
- Profiling implementation: `Function/Clean/data_profile.py`
- Winsorization implementation: `Function/Clean/winsorize.py`