|
|
--- |
|
|
license: mit |
|
|
language: en |
|
|
--- |
|
|
|
|
|
# Bitcoin Price Prediction Dataset (60d History -> 4d Forecast) |
|
|
|
|
|
This dataset is designed for fine-tuning Large Language Models (LLMs) on a time-series forecasting task. Each sample contains 60 days of historical financial and social media data to predict the next 4 days of Bitcoin's closing price. |
|
|
|
|
|
## Dataset Description |
|
|
|
|
|
The dataset was generated by combining two primary sources: |
|
|
1. **Financial Data:** Historical daily prices for Bitcoin (BTC), Gold (GC=F), Crude Oil (CL=F), S&P 500 (^GSPC), and the US Dollar Index (DX-Y.NYB) were fetched from Yahoo Finance. |
|
|
2. **Social Media Data:** The [Bitcoin Tweets Dataset](https://www.kaggle.com/datasets/kaushiksuresh147/bitcoin-tweets) was used to extract daily tweet volume and sample tweet texts. |
|
|
|
|
|
### Features Included in the Prompt: |
|
|
- **60-Day Historical BTC Prices**: The core time-series data. |
|
|
- **Technical Indicators**: 14-day RSI, 50-day EMA, and 200-day EMA for Bitcoin. |
|
|
- **Macroeconomic Context**: Daily closing prices of Gold, Oil, S&P 500, and the US Dollar Index. |
|
|
- **Social Media Sentiment**: A sample of 4 tweets for the most recent day in the historical window. |
|
|
|
|
|
## Data Fields |
|
|
|
|
|
The dataset is structured for instruction fine-tuning and contains three columns: |
|
|
|
|
|
- `instruction`: A string containing the comma-separated closing prices of Bitcoin for the last 60 days. |
|
|
- `input`: A detailed text block containing all the contextual information (technical indicators, macro data, sample tweets) for the most recent day of the historical period. |
|
|
- `output`: A string containing the comma-separated closing prices of Bitcoin for the next 4 days (the prediction target). |
|
|
|
|
|
## Splits |
|
|
|
|
|
- **train**: The training split (`train_dataset.json`). |
|
|
- **test**: The validation/test split (`val_dataset.json`). |
|
|
|
|
|
## Intended Use |
|
|
|
|
|
This dataset is intended to be used with the `SFTTrainer` from the TRL library or similar frameworks to fine-tune models like Llama 3 for complex, multi-modal time-series forecasting tasks. |
|
|
|
|
|
Example prompt structure: |
|
|
```json |
|
|
{{ |
|
|
"instruction": "45123.45, 45321.89, ... (60 days of prices)", |
|
|
"input": "Based on the historical data from the last 60 days, predict the Bitcoin closing prices for the next 4 days. The prediction period starts on 2023-11-20 (Weekday). The most recent day's technical analysis (2023-11-19): ...", |
|
|
"output": "46123.78, 46050.12, 46300.50, 45987.32" |
|
|
}} |
|
|
``` |
|
|
|
