TradingAgents/proposals/sentiment_divergence_fix.md

3.1 KiB

🐛 The Problem: The "Sentiment-Price Divergence" Blind Spot

Currently, TradingAgents derives signals primarily from SentimentAnalysisAgent and NewsDataAgent using a linear assumption:

  • Positive News/Sentiment → Bullish Signal
  • Negative News/Sentiment → Bearish Signal

However, market structures (especially in momentum markets like A-shares or Crypto) are non-linear. This logic fails to capture "Narrative Exhaustion" and the "Consensus Trap".

The "Bug" in Current Logic

  1. The "Consensus Trap" (Sell the News): When News Sentiment reaches 100 (Peak Consensus), the model treats it as a strong Buy. In reality, this often indicates Distribution (Smart Money selling to Retail on good news). The model misses the turning point where "Obsession" becomes a rigid trap.
  2. Static vs. Dynamic Sentiment: The agents treat sentiment as a static data point. They miss the evolution of market psychology: Event → Obsession Formation → Peak Consensus → Breakdown. A "Positive" sentiment at the start of a trend means something completely different than a "Positive" sentiment at the top.
  3. Ignoring "Path of Least Resistance": As Jesse Livermore noted, prices move in the line of least resistance (determined by Price/Volume), not by headlines. The model often overweights News, leading to false signals when price action contradicts the narrative (e.g., Price drops despite Good News).

💡 The Proposal: Integrating "Market Psychology" & Divergence Checks

I propose enhancing the RiskAgent to act as a Reality Check against raw sentiment, incorporating basic market methodology.

1. Sentiment Divergence Check (The "Livermore" Logic)

  • Divergence Bearish: If News Sentiment is High but Price Action is Bearish, override the signal to Strong Sell.
    • Why: This indicates the "Path of Least Resistance" is Down, despite the noise. Smart money is exiting.
  • Divergence Bullish: If News Sentiment is Low but Price Action is Bullish, override to Strong Buy.
    • Why: Accumulation often happens in fear.

2. Obsession Phase Detection (The "Consensus" Logic)

  • Introduce a concept of "Narrative Saturation".
  • If media coverage is Uniform and Extremely Positive (e.g., >90% positive), flag this as High Risk of Reversal, even if the score is perfect.
  • Actionable Advice: The system should warn: "Consensus is peaking. If price fails to break resistance, exit immediately."

3. Priority: Price > News

  • Rule: News explains why the move happened, but Price tells us what is happening.
  • The model should prioritize the Path of Least Resistance (Price/Volume) over the loudest narrative. News should only be used as a confirmation, not a primary trigger.

📉 Impact

By addressing this, TradingAgents will move from a "Data Summarizer" to a true "Trading Decision Engine". It will help users avoid "Bull Traps" caused by blindly following positive news at market tops and understand that sentiment is often a contrarian indicator at extremes.