Sentiment Methodology V2

Effective 2026-02-16T05:14:00Z, the sentiment scoring pipeline has been upgraded to Methodology V2. This is Phase 1 of a two-phase deployment.

What Changed

Both AI models in the sentiment scoring system have been replaced with newly trained versions. The scoring architecture remains the same - a primary model provides the default classification, while a secondary model provides confidence-based quality gating. Only the underlying models have changed.

Key Improvements

  • Negative sentiment detection improved by approximately 50%. The previous models missed a significant portion of bearish tweets, systematically underreporting negative sentiment. The new models correct this bias.
  • Model confidence quality nearly doubled. The secondary model’s confidence scores are now substantially more reliable, improving the accuracy of neutral and override decisions.
  • Training data expanded by approximately 3.5x. The new models were trained on a much larger and more balanced dataset, with particular focus on improving coverage of negative examples.

Impact on Data

All data produced after the cutover timestamp reflects V2 methodology:

  • sentiment_model_version changed from v1.0 to v2.0
  • methodology_regime field added with value v2
  • Negative sentiment ratios will be noticeably higher (correcting previous underreporting)
  • Overall sentiment scores will shift closer to neutral (less positive bias)
  • Post volume is unchanged in Phase 1

Cutover Details

DetailValue
Cutover timestamp2026-02-16T05:14:00Z
Version identifierv2.0
Regime fieldmethodology_regime: "v2"

Historical data (since December 2025) remains available but should not be directly compared with V2 data without accounting for the methodology change. The methodology_regime field enables programmatic filtering.

Phase 2 (Pending)

A crypto relevance filter will be deployed in Phase 2, removing non-crypto noise before sentiment scoring. This will improve sentiment signal quality by ensuring only relevant posts are included in aggregated metrics.


Cutover timestamp: 2026-02-16T05:14:00Z.