← Back to Skills Marketplace
gechengling

Financial Engineer Digital Employee

by lingfeng-19 · GitHub ↗ · v1.0.0 · MIT-0
cross-platform ⚠ suspicious
44
Downloads
0
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install financial-engineer-digital-employee
Description
金融风控建模数字员工——覆盖数据探查、单变量分析、特征工程、LR评分卡、XGBoost/DNN建模、超参数调优、模型解释、多模型对比、分群建模、DeepModel集成全流程。从数据到模型上线的一站式机器学习建模能力。
README (SKILL.md)

\r \r

Financial Engineer / 金融工程专家数字员工\r

\r

⚠️ SECURITY NOTICE / 安全声明\r

  • Type: Educational reference / analytical framework ONLY\r
  • No executable code, scripts, or binaries are included in this skill\r
  • No persistent storage, network calls, background execution, or credential collection\r
  • All outputs are for reference only and require human review before real-world application\r
  • This skill does NOT provide financial, legal, or insurance advice\r
  • Users must exercise their own judgment and consult qualified professionals\r \r

Overview\r

\r 金融工程专家数字员工,集成以下14项核心能力:\r \r

  1. Auto Experiment — 【自主实验循环·特征组级探索】当用户说"自动探索哪些特征组有用"、"批量对比不同特征组贡献"、"帮我自主试多种方案"、"自动跑一下"、"试各种特征组合"、"自动特征选择"、"消融分析"时使用。核心能力\r
  2. Data Profiling — 【数据轮廓速览】当用户刚上传数据、问"数据有多少行"、"有哪些字段"、"数据长什么样"时使用。描述数据结构(行列数、字段类型、缺失率、分布特征、数据质量),不计算建模指标。无需目标变量。与 featu\r
  3. Dnn Modeling — 【DNN深度学习建模】当用户需要"深度学习"、"DNN建模"、"神经网络"、"MLP"、"跑个DNN"、"深度模型"时使用。基于 PyTorch 实现多层全连接网络(MLP)进行二分类建模。支持 Ba\r
  4. Dnn Tuning — 【DNN调参】当用户说"DNN调参"、"深度学习调参"、"调整网络结构"、"调整dropout"、"DNN过拟合"时使用。基于 Optuna TPE 贝叶斯优化,搜索网络架构(层数、宽度、dropou\r
  5. Feature Analysis — 【特征深度分析】当用户需要"全面分析特征"、"生成特征报告"、"看相关性"、"推荐建模方案"、"跑一遍特征"、"特征全景"、"哪些特征有用"时使用。包含四大维度:基础统计、IV值、PSI稳定性、相关性\r
  6. Lr Modeling — 【LR评分卡建模】当用户需要"逻辑回归"、"LR建模"、"评分卡"、"scorecard"、"WoE建模"、"跑个LR"时使用。基于 WoE 编码 + Logistic Regression 训练标准\r
  7. Lr Tuning — 【LR评分卡调参】当用户说"LR调参"、"评分卡调优"、"调整分箱数"、"调整C值"、"LR过拟合"时使用。基于 Optuna TPE 贝叶斯优化,联合搜索 WoE 分箱参数(max_n_bins、i\r
  8. Model Comparison — 【多模型效果对比】当用户需要"对比算法"、"多模型对比"、"XGB和LR对比"、"三种算法都试试"、"哪个算法好"、"都跑一下对比看看"、"自动调参后对比"、"公平对比"时使用。基于同一份数据切分,横\r
  9. Model Explanation — 【模型解释】当用户问"为什么模型这样预测"、"哪些特征最重要"、"解释一下这个样本"、"模型怎么做的决策"、"feature importance"、"为什么给这个分数"时使用。基于SHAP的模型解释\r
  10. Segment Modeling — 【分群自主探索】当用户需要"分群建模"、"客群拆分训练"、"探索最优分群方案"、"按xxx分组建模"、"不同客群分别训练"、"分群策略"时使用。支持规则分群(人类先验)、聚类分群(无监督发现)、决策树\r
  11. Univariate Analysis — 【单变量分析】当用户提到"分析某个/某些特征"、"看看xxx特征"、"xxx特征分布"、"帮我分箱"、"交叉分布"、"筛选特征"、"分析一下xxx"等涉及少量特征分析的请求时使用。支持两种模式:(1)\r
  12. Xgb Deepmodel — 【DeepModel深度建模】当用户需要"分群集成建模"、"分群+Stacking融合"、"子模型训练+集成"时使用。核心能力:按业务分群条件训练多个 XGBoost 子模型,通过 Stacking \r
  13. Xgb Modeling — 【XGBoost建模】当用户需要"训练模型"、"跑一个基线"、"建个模型试试"、"对比特征方案效果"、"预测一下"、"跑个xgb"、"看AUC/KS/BCR"时使用。支持多特征方案对比、AUC/KS/\r
  14. Xgb Tuning — 【XGBoost超参数调优 — 唯一调参入口】当用户说"帮我调参"、"模型过拟合了怎么办"、"调整learning_rate/max_depth等参数"时使用。核心能力:基于 Optuna TPE 贝\r \r ---\r \r

Module 1: Auto Experiment\r

\r

Phase 1: 特征组独立评估\r

  • 自动发现数据中的特征分组(按前缀聚合: firefly_*, mob3_*, umeng_* 等)\r
  • 逐组单独建模,评估每组的独立预测能力\r
  • 输出各组的独立 AUC/KS 排名\r \r

Module 2: Data Profiling\r

\r | 参数 | 必选 | 默认值 | 说明 |\r |------|:----:|--------|------|\r | --data_path | ✅ | - | 数据文件路径(parquet/csv) |\r | --output | | data_profiling_report.md | 报告输出路径 |\r | --output_dir | | ./outputs/\x3Cts> | 产物输出目录 |\r | --output_name | | data_profiling_report | 报告基名(不含扩展名) |\r | --config | | - | JSON 配置文件路\r \r

Module 3: Dnn Modeling\r

\r 基于 PyTorch 实现 MLP (Multi-Layer Perceptron) 进行二分类建模。\r \r

Module 4: Dnn Tuning\r

\r

执行模式\r

\r | 模式 | 触发条件 | 行为 |\r |------|---------|------|\r | 交互式(默认) | 用户说"调参"/"帮我调一\r \r | 模式 | 触发条件 | 行为 |\r |------|---------|------|\r | 交互式(默认) | 用户说"调参"/"帮我调一下DNN" | 每轮暂停等待用户反馈 |\r | AUTO | 用户说"自动调优"/"帮我调到最优" | Agent 自动迭代直到收敛 |\r \r 默认模式: 交互式(更安全,用户可控)\r \r ---\r \r

Module 5: Feature Analysis\r

\r 基于 scripts/analyzer.py 主脚本,对数据集特征进行全面分析并生成 Markdown 报告。\r \r

Module 6: Lr Modeling\r

\r 基于 WoE (Weight of Evidence) 编码 + Logistic Regression 进行二分类评分卡建模。\r \r

Module 7: Lr Tuning\r

\r

执行模式\r

\r | 模式 | 触发条件 | 行为 |\r |------|---------|------|\r | 交互式(默认) | 用户说"调参"/"帮我调一\r \r | 模式 | 触发条件 | 行为 |\r |------|---------|------|\r | 交互式(默认) | 用户说"调参"/"帮我调一下LR" | 每轮暂停等待用户反馈 |\r | AUTO | 用户说"自动调优"/"帮我调到最优" | Agent 自动迭代直到收敛 |\r \r ---\r \r

Module 8: Model Comparison\r

\r 设计原则(v3 重要变更):\r

  • 只陈述事实,不做主观推荐:删除了旧版的"五维加权综合评分"与"门禁淘汰"。主观权重本质上将多个不可比指标压成一个数,丢失信息并引入拍脑袋假设;门禁阈值则不同业务差异巨大。\r
  • Pareto 前沿(无主观权重):在 OOT AUC / OOT KS / BCR@10% / KS Gap / PSI 五个方向明确的目标上,识别未被任何算法严格优于的候选集。\r
  • LLM 推理接手:报告末尾提供「已知事实清单 + 开放问题」,交由对话中的 AI 结合业务背景推理取舍。\r
  • 同数据同切分 → 消除数据差异对对比结论的干扰\r
  • 多维指标原始\r \r

Module 9: Model Explanation\r

\r 基于 SHAP (SHapley Additive exPlanations) 对 XGBoost 模型进行可解释性分析,生成包含可视化图表的 Markdown 报告。\r \r

Module 10: Segment Modeling\r

\r

探索空间(策略 × 参数 × 组合)\r
        │\r
        ▼\r
Try → Measure → Keep/Discard → Repeat\r
        │\r
        ▼\r
    最优分群方案 + 子模型\r
```\r
\r
## Module 11: Univariate Analysis\r
\r
基于 `scripts/analyzer.py` 主脚本,对指定特征进行单变量级别的分布分析或预测力评估。\r
\r
## Module 12: Xgb Deepmodel\r
\r
> 参数 spec:`_vendor/xgb_cli.py` 中 `domain=deepmodel-sub`。复杂嵌套 JSON(segments / features_per_segment / pos_weight_per_segment)**必须**走 `--config`。\r
\r
## Module 13: Xgb Modeling\r
\r
基于 XGBoost 进行二分类建模,支持自动特征筛选、多方案对比、稳定性分析。可运行在任何 Python 环境,无平台耦合。调参请用 [`xgb-tuning`](./xgb-tuning/SKILL.md)。\r
\r
## Module 14: Xgb Tuning\r
\r
### 执行模式\r
\r
| 模式 | 触发条件 | 行为 |\r
|------|---------|------|\r
| **交互式**(默认) | 用户说"调参"/"帮我调一\r
\r
| 模式 | 触发条件 | 行为 |\r
|------|---------|------|\r
| **交互式**(默认) | 用户说"调参"/"帮我调一下"/"优化一下" | 每轮暂停等待用户反馈 |\r
| **AUTO** | 用户说"自动调优"/"帮我调到最优"/"一直调到收敛" | Agent 自动迭代直到收敛,每轮输出进度 |\r
\r
**默认模式**: 交互式(更安全,用户可控)\r
\r
---\r
\r
## Disclaimer / 免责声明\r
\r
> ⚠️ **重要声明**\r
> - 本技能提供参考框架和分析建议,不构成任何形式的投资建议、法律意见或专业判断\r
> - 所有分析结果仅供参考,最终决策须由具备相应资质的专业人员作出\r
> - AI生成的分析不代表任何机构的官方立场或承诺\r
> - 用户应结合实际情况独立判断
Usage Guidance
Review before installing. Treat this as an operational financial modeling workflow, not just static documentation: confirm before running scripts, use only intended datasets, choose output directories deliberately, and inspect or delete generated reports and model artifacts that may contain sensitive derived information.
Capability Assessment
Purpose & Capability
The stated purpose, financial risk modeling and model analysis, matches most modules, including profiling, feature analysis, scorecards, XGBoost/DNN modeling, tuning, explanation, and segment modeling.
Instruction Scope
The manifest and security notice frame the skill as advisory and no-executable-code, while the body gives operational instructions for scripts, CLI parameters, automated training, tuning loops, and model comparison. No scripts are packaged, but the instructions could still drive an agent to execute local workflows if matching files exist.
Install Mechanism
The package contains only a single markdown skill file, declares no allowed tools, has no dependencies, and both static scan and VirusTotal telemetry are clean.
Credentials
Financial modeling can involve sensitive user datasets, but the artifact does not request credentials, network access, broad filesystem access, or account mutation. Users should scope any supplied data and outputs carefully.
Persistence & Privilege
The security notice says there is no persistent storage, but the body defines report output paths, output directories, Markdown reports, visualizations, and model-result artifacts. That storage is purpose-aligned but under-disclosed.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install financial-engineer-digital-employee
  3. After installation, invoke the skill by name or use /financial-engineer-digital-employee
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.0
- Initial release of "Financial Engineer" digital employee skill (v1.0.0). - Provides a comprehensive end-to-end machine learning workflow for financial risk modeling, covering 14 core modules (data profiling, feature analysis, LR/scorecard modeling, XGBoost/DNN models, model tuning, model interpretation, group modeling, and more). - All functionality is advisory and reference-only; no executable code or persistent storage is included. - Outputs require human review and do not constitute financial or legal advice. - Includes detailed security and disclaimer notices to ensure safe and responsible usage.
Metadata
Slug financial-engineer-digital-employee
Version 1.0.0
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 1
Frequently Asked Questions

What is Financial Engineer Digital Employee?

金融风控建模数字员工——覆盖数据探查、单变量分析、特征工程、LR评分卡、XGBoost/DNN建模、超参数调优、模型解释、多模型对比、分群建模、DeepModel集成全流程。从数据到模型上线的一站式机器学习建模能力。 It is an AI Agent Skill for Claude Code / OpenClaw, with 44 downloads so far.

How do I install Financial Engineer Digital Employee?

Run "/install financial-engineer-digital-employee" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Financial Engineer Digital Employee free?

Yes, Financial Engineer Digital Employee is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Financial Engineer Digital Employee support?

Financial Engineer Digital Employee is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Financial Engineer Digital Employee?

It is built and maintained by lingfeng-19 (@gechengling); the current version is v1.0.0.

💬 Comments