| | --- |
| | license: mit |
| | tags: |
| | - time-series |
| | - crypto |
| | - forecasting |
| | - transformer |
| | --- |
| | |
| | # crypto_volatility_prediction_informer |
| | |
| | ## Overview |
| | This model implements the Informer architecture for long-sequence time-series forecasting. It is specifically tuned to predict Bitcoin (BTC) and Ethereum (ETH) price volatility over a 24-hour horizon based on a rolling 7-day window of hourly OHLCV data. |
| | |
| | |
| | |
| | ## Model Architecture |
| | The Informer model addresses the $O(L^2)$ complexity of standard Transformers using: |
| | - **ProbSparse Self-Attention**: Reduces complexity to $O(L \log L)$. |
| | - **Self-attention Distilling**: Highlights dominant features across temporal dimensions. |
| | - **Generative Decoder**: Predicts long-term sequences in one forward step to prevent error accumulation. |
| | |
| | ## Intended Use |
| | - **Risk Management**: Estimating potential volatility spikes for automated trading desks. |
| | - **Portfolio Hedging**: Generating signals for derivative positioning. |
| | - **Market Research**: Analyzing the temporal dependencies of crypto-asset fluctuations. |
| | |
| | ## Limitations |
| | - **Black Swan Events**: The model cannot predict volatility caused by external regulatory news or exchange failures not reflected in historical price patterns. |
| | - **Non-Stationarity**: Crypto markets are highly non-stationary; the model requires frequent re-training (e.g., every 48 hours). |
| | - **Financial Risk**: This tool is for informational purposes and is not financial advice. |
