Advantages of Large Language Models
Traditional time series models (such as ARIMA and LSTM) have certain capabilities in processing numerical price data, but they struggle to effectively utilize text information. The emergence of large language models has changed this situation:
Multimodal Understanding Ability
Large language models can process both numerical and text data simultaneously, converting text information from social media posts, news articles, and forum discussions into quantifiable features.
Contextual Understanding
Unlike simple keyword counting, large language models can understand the context and semantics of text. For example:
- The emotional polarity of "Bitcoin skyrockets" and "Bitcoin plummets" is completely different
- Sarcasm and irony can be identified
- Professional terms and slang can be correctly understood
Long Text Processing
Large language models can handle long documents, capture key information, generate summaries and sentiment scores, and provide rich feature inputs for prediction models.
Project Technical Scheme
Data Collection Layer
The project needs to collect multi-source data:
Price data: Obtain historical price data from cryptocurrency exchange APIs, including opening price, closing price, highest price, lowest price, trading volume, etc.
Social media data: Obtain relevant posts and comments through Twitter API, Reddit API, etc., including text content and interaction data (likes, retweets, comment counts).
News data: Crawl headline news from cryptocurrency news websites and extract titles and summaries.
Search trend data: Obtain search popularity data through Google Trends API.
Feature Engineering Layer
Text feature extraction:
- Use large language models to perform sentiment analysis on social media posts and news
- Extract topics and keywords
- Generate text embedding vectors
- Calculate discussion popularity and spread speed
Numerical feature construction:
- Technical indicators: Moving averages, RSI, MACD, etc.
- Volatility indicators
- Trading volume changes
Prediction Model Layer
The project may adopt the following model architectures:
Multimodal fusion model: Fuse text features and numerical features and input them into the prediction model.
Time series model: Use Transformer or LSTM to handle time series dependencies.
Ensemble method: Combine prediction results from multiple models to improve robustness.
Prediction Objectives
Since it is short-term prediction, the project may focus on:
- Price direction (up/down/flat) in the next hour/day
- Price fluctuation range
- Trading volume prediction
