factor-mining - SKILL.md Agent Skill

name: factor-mining shortcut: factor description: Guide user through factor research and mining. Trigger when user asks to "find effective factors", "mine factors", "research alpha factors", "help me design a factor", "factor analysis", or similar requests about discovering new trading factors. description_zh: 引导用户进行因子研究与挖掘，包括假设生成、表达式构建、有效性评估和因子保存。

Factor Mining Workflow

Guide the user through discovering, testing, and saving effective trading factors.

Pre-requisites (MUST confirm)

Which exchange to analyze (Hyperliquid or Binance)
Which symbol(s) to focus on (e.g., BTC, ETH)
What trading style or hypothesis they have in mind (optional)

Workflow

Phase 1: Survey Existing Factors

Use query_factors to check what's already computed:

Show top factors ranked by |ICIR| for the target symbol
Highlight factors with |ICIR| > 1.5 (strong) or |IC| > 0.05 (meaningful)
Note which categories are well-covered vs under-explored

[CHECKPOINT] Present existing factor landscape. Ask user:

Any patterns they notice?
Which direction to explore (momentum, volatility, microstructure, custom)?

Phase 2: Hypothesis Generation

IMPORTANT: Call get_factor_functions first to get the full list of available functions and their signatures. Do NOT guess function names or signatures — always check the registry.

Based on user's interest, generate 2-3 factor hypotheses:

Approach A: Expression-based (no web search needed)

Combine existing indicators in new ways
Common patterns: ratio (EMA7/EMA21-1), acceleration (ROC3-ROC10), normalized deviation ((close-SMA20)/STDDEV(close,20))
Cross-category combinations (momentum + volatility, trend + volume)

Approach B: Research-driven (use web_search + fetch_url)

If user wants inspiration from academic research, known factor libraries, or quant blogs
Step 1: Search for sources — prioritize academic and code repositories:
- For known factor sets (e.g., WorldQuant 101 Alphas): site:github.com WorldQuant alpha101 formula or site:arxiv.org 101 Formulaic Alphas
- For specific factor numbers: site:github.com "Alpha#101" formula
- For general quant research: site:arxiv.org cryptocurrency momentum factor
Step 2: Fetch full content — use fetch_url on the most promising URL to read the actual formula/paper
Step 3: Translate to expression — convert the retrieved formula into a factor expression compatible with our system
Do NOT repeatedly search with different keywords hoping snippets contain the answer

[CHECKPOINT] Present hypotheses with rationale. Let user pick which to test.

Phase 3: Test & Evaluate

For each chosen hypothesis:

Use evaluate_factor with the expression + target symbol
Interpret results:
- ICIR > 2.0: Very strong predictive power
- ICIR 1.0-2.0: Moderate, worth exploring
- ICIR < 0.5: Weak, likely noise
- Win rate > 55%: Directionally useful
- Check across forward periods (1h/4h/12h/24h) for decay pattern
Compare with existing built-in factors — is the new factor adding value?

[CHECKPOINT] Present evaluation results in a comparison table. Recommend which factors to keep.

Phase 4: Save & Compute

For factors that pass evaluation:

Use save_factor with a descriptive name and clear description
Use compute_factor to run full evaluation across all watchlist symbols
Suggest the user visit Factor Library to view results

[CHECKPOINT] Summarize what was saved. Suggest next steps:

Test more variations of successful factors
Consider how to integrate into trading strategy (Phase 3-4 of factor system)
Set up periodic re-evaluation

Tips for the AI

Always call get_factor_functions before writing any expression
Always use English for web search queries (better results)
When comparing factors, use a markdown table for clarity
Explain IC/ICIR/win_rate in plain language for less experienced users
If a factor has high IC but low ICIR, explain it means inconsistent signal
Suggest testing both the factor and its negation (multiply by -1)