quantskills/skill-pandaai-factor-online

Chinese Simplified | English

Solidify the hands-on steps and pitfalls of the PandaAI Factor Competition into a skill, allowing any AI tool to take you from an empty machine to a submittable factor.

---

📖 What is this?

A portable Agent skill, through pandaai-cli on PandaAI platform mines quantitative factors. Compatible with Claude Code, Cursor, Codex, Kimi Code, Gemini CLI, and any agent that reads AGENTS.md.

There are several hurdles where PandaAI is not found in documentation but actually consumes time and computing power:

• On a new machine, pandaai-cli login will inevitably fail: the CLI loads the configuration file before distributing subcommands, and creates the file It is precisely the login command itself.
• MEAN(CLOSE, 20) is not the 20-day moving average. It can parse, operate, and return a factor that looks reasonable but actually measures something else.
• Group numbers are fixed in ascending order by factor values, with the multi-head end determined by --factor-direction rather than following the number. If you misread one end, all conclusions will be reversed.
• Platform headlines are long-short annualized, implying short positions that A-share participants cannot build; Turnover rate is listed separately and does not include profits.
• The backtest window for CLI and server may differ; Before long samples go live, the current server capability must be probed first. Upgraded to CLI 0.1.6 on 2026-08-10, The stock pool is fixed at all A stocks in the Shanghai and Shenzhen sectors; The backtest duration is still based on actual server-side detection results, and out-of-sample validation still requires explicitly establishing a separate research object.

The skill includes these conclusions, 348 fields and 137 operators with complete references, as well as a set of research workflows, Prevents Agents from spending all their computing power on hundreds of variants of the same idea.

🚀 Get Started Quickly

One-Sentence Installation (Recommended)

Paste this sentence into Claude Code, Cursor, Codex, or Kimi Code:

No additional steps are needed. After the agent clone, it reads the root directory of the repository AGENTS.md. Three things to do during a cold start: confirm the repository is located in the persistent directory, Run the installation script and follow SKILL’s core process. The core process is an explicit contract, so different tools perform consistently.

It first performs a zero-computing power check, filling gaps one by one—Python version, CLI installation, account registration, competition registration, account password— Then guide you to log in. The password is always entered by you on the terminal, and the agent does not touch you. After the physical exam is all green, it converts the computing power balance into the number of runs allowed. We will confirm the rebalancing cycle, backtest range, and current budget with you, and finally provide a list of candidates for your review. **No hash power is spent before approval. **

Manual installation

The prerequisite is only Python 3.10 or later. If there is no or outdated version, uv is the shortest path and does not depend on Python itself:

You can also use UV: Python can be used from python.org, [T7]],