Lesson Learned #093: Google Recommender CAV Not Useful for Trading
Date: January 7, 2026
Severity: INFO
Category: Research, Strategy Evaluation
Summary
Evaluated Google's Recommender System breakthrough using Concept Activation Vectors (CAVs) for detecting semantic intent. Determined it is NOT useful for our trading system.
Technical Analysis
What Google's CAV Does
- Uses Concept Activation Vectors to interpret USERS (not models)
- Translates subjective "soft attributes" (funny, cute, boring) into vectors
- Personalizes content recommendations (YouTube, Google Discover)
- Tested on MovieLens20M dataset
Why It's NOT Useful for Trading
| Google's CAV | Our Trading System |
|---|---|
| Problem: "What does 'funny' mean to THIS user?" | Problem: "What is the CROWD saying about SPY?" |
| Goal: Personalize content to individuals | Goal: Aggregate market sentiment |
| Data: User interaction history | Data: Public posts, news |
| Output: Personalized recommendations | Output: Buy/Sell/Hold signals |
Critical Mismatch
CAVs solve personalization. We need aggregation.
For market sentiment, we don't care if one user thinks "bullish" means slightly optimistic vs extremely positive. We care about volume and direction of crowd sentiment.
Decision
DO NOT IMPLEMENT - Would add:
- Unnecessary complexity
- More failure points
- No material improvement to trading signals
- Violates CLAUDE.md: "100% operational security"
What We Have (Appropriate)
-
src/utils/unified_sentiment.py- Multi-source weighted aggregation - Keyword-based sentiment with News (40%), Reddit (35%), YouTube (25%)
- This IS the right approach for market sentiment analysis
CEO Validation
- CEO asked for honest assessment
- Answered: "This is FLUFF for our use case"
- CEO accepted the analysis
Tags
research, google_cav, recommender_system, sentiment_analysis, strategy_evaluation, not_implemented
This lesson was auto-published from our AI Trading repository.
More lessons: rag_knowledge/lessons_learned
Top comments (0)