Leveraging Large Language Models for Sentiment Analysis: Multi-Modal Analysis of Decentraland's MANA Token

Battle Count: 107

The paper addresses cryptocurrency price prediction using sentiment and multi-modal features, which is directly relevant to quantitative trading in crypto markets. However, the practical relevance is limited by: (1) negative R² values indicating poor directional prediction, (2) very small dataset (366 days), (3) focus on a single niche token (MANA) in a virtual world economy, (4) sentiment reflecting platform UX rather than market fundamentals, and (5) no trading strategy backtesting or transaction cost analysis. The methodology (LLM sentiment + LSTM) is transferable to more liquid crypto assets, but the specific findings have limited direct applicability to production trading systems.

Moderate complexity. The pipeline involves: (1) Discord data scraping via DiscordChatExporter, (2) RoBERTa inference for sentiment classification (straightforward via Hugging Face), (3) daily sentiment aggregation with confidence weighting, (4) LSTM model training with standard PyTorch/Keras. The main complexity lies in data preprocessing, sentiment aggregation logic, and ensuring proper temporal alignment between sentiment scores and price data. No custom model architectures are proposed; all components use existing tools. The 8-fold cross-validation adds computational overhead but is standard.

All data and code are publicly accessible on GitHub. The paper uses well-known pre-trained models (RoBERTa from Hugging Face), standard LSTM architectures, and publicly available financial data from CoinMarketCap. Discord data collection method (DiscordChatExporter) is specified. However, the exact LSTM hyperparameters (number of layers, hidden units, learning rate, epochs) are not fully detailed in the main text. The 8-fold cross-validation results are provided in the appendix.