BERT-based model predicts MANA token prices using Discord sentiment
Community chat signals beat price-only models by significant margin...
A team of researchers led by Xintong Wu and Luyao Zhang (arXiv 2605.20192) has demonstrated that large language models can extract actionable trading signals from community platforms. They applied a BERT-based sentiment analysis model to Decentraland's Discord server, classifying messages as positive, negative, or neutral. The overwhelmingly neutral-but-positive-skew sentiment was then combined with trading volume and market cap data to feed an LSTM (long short-term memory) neural network designed for cryptocurrency price prediction.
The multi-modal LSTM—trained on sentiment scores plus traditional financial features—significantly outperformed a baseline LSTM that relied solely on historical MANA token prices. The paper specifically studies Decentraland, a blockchain-based virtual world where MANA is used for land purchases, governance, and in-world transactions. This work suggests that community-derived signals, when properly processed with modern NLP, can enhance forecasting for Metaverse-native assets. The findings open the door for future research combining immersive environments, NLP, and crypto market analysis.
- BERT-based LLM classified Discord messages to extract sentiment scores for Decentraland's MANA token community.
- Multi-modal LSTM combining sentiment, volume, and market cap outperformed price-only baseline in return forecasting.
- Discord sentiment was predominantly neutral with a positive skew, yet still provided predictive value beyond price history.
Why It Matters
Community sentiment analysis now quantifiably improves crypto price predictions, practical for Metaverse asset traders and DAO analysts.