Alpha Zoo

Purpose

When the user asks about prebuilt cross-sectional alphas — Kakushadze 101, GTJA 191, Qlib 158, Fama-French / Carhart — or wants to bench a whole zoo on an investable universe (CSI 300, S&P 500, BTC-USDT, ...), this skill orients you. The zoo is the curated library; the bench is the evaluator.

Tools Available

Tool	When to use
alpha_zoo	Browse the library. action=list_alphas to enumerate (filterable by zoo / theme / universe), action=get_alpha for one alpha's metadata, action=health for registry load status.
alpha_bench	Run IC / IR on one alpha or a whole zoo over a universe + period. Emits an HTML report.
factor_analysis	Ad-hoc factor evaluation from a user-supplied factor CSV + return CSV. Use this when the user has their own factor (not in the zoo).

Decision Tree

• "list all momentum alphas" → alpha_zoo with action=list_alphas, theme=momentum.
• "show me gtja191_alpha_001" → alpha_zoo with action=get_alpha, alpha_id=gtja191_alpha_001.
• "bench all of GTJA 191 on CSI 300 from 2020 to 2024" → alpha_bench with zoo=gtja191, universe=csi300, period=2020-2024.
• "is the registry healthy" → alpha_zoo with action=health — surfaces loaded, failed, and per-error reasons.
• User uploads my_factor.csv → factor_analysis (zoo tools are for prebuilt alphas only).

Zoo Inventory

Zoo	Description	Approx. count
kakushadze101	Formulaic alphas from Kakushadze's 2015 paper. Mix of momentum, reversal, volume, and microstructure.	~101
gtja191	Guotai Junan 191 alphas — A-share focused cross-sectional factors.	~191
qlib158	Microsoft Qlib's 158 alpha factors — features tuned for ML pipelines.	~158
classical	Fama-French 3/5-factor + Carhart momentum.	<10

Counts are nominal; check alpha_zoo action=health for the live count currently loaded.

Constraints

• No per-stock per-date factor values are surfaced to the agent. IC results are aggregate stats (mean / std / IR / positive-ratio); the HTML report shows top-N by IR plus formulas, never the underlying panel.
• Lookahead is banned in the operator set. delta(df, d) requires d >= 1; the negative-shift Ref(df, -n) form does not exist. See docs/alpha-zoo/spec.md for the full operator catalogue.
• Universe loaders may not be wired for every market yet. When alpha_bench returns universe loader for X not yet implemented, that's the W2 scaffold — the universe is recognised but the data pull lands in W4.
• Do not expose absolute filesystem paths in agent output. The bench tool writes to ~/.vibe-trading/reports/ by default; refer to it by that shorthand, not by the resolved absolute path.
• alpha_zoo is read-only. alpha_bench writes a single HTML file per run — no scratch state elsewhere.

Common Pitfalls

• Filter mismatch on list_alphas: theme / universe must match the alpha's declared metadata exactly (e.g. equity_cn, not cn or china).
• Calling alpha_bench with both alpha_id and zoo set — they are mutually exclusive; pick one.
• Empty registry (loaded=0) means no zoo modules are populated yet; treat it as "zoos pending W3 porting" rather than a bug.

Reference

• Operator catalogue: docs/alpha-zoo/spec.md
• Registry contract: src/factors/registry.py (frozen; do not modify)
• IC / layered NAV math: src/factors/factor_analysis_core.py