metadata
license: mit
tags:
- time-series
- forecasting
- finance
- crypto
crypto_volatility_forecaster
Overview
This model utilizes a Time-Series Transformer architecture to predict the volatility of major cryptocurrencies (e.g., BTC, ETH). By processing historical price action and volume data, it forecasts a probabilistic distribution of future price movements over a 24-hour window based on a 7-day look-back period.
Model Architecture
The model implements a specialized Encoder-Decoder Transformer designed for sequential numerical data.
- Temporal Embedding: Captures hourly and daily seasonalities.
- Distribution Head: Instead of point forecasts, it outputs parameters for a Student's t-distribution, which is better suited for the "fat tails" observed in financial market data.
- Context Window ($L$): 168 hours.
- Prediction Horizon ($H$): 24 hours.
The loss function used is the Negative Log-Likelihood ($NLL$):
Intended Use
- Risk Management: Estimating potential Value at Risk (VaR) for digital asset portfolios.
- Algorithmic Trading: Providing volatility signals as features for automated execution strategies.
- Financial Research: Studying market regime shifts and anomaly detection.
Limitations
- Black Swan Events: Cannot predict volatility spikes caused by external shocks (e.g., regulatory changes, exchange failures) not present in historical price data.
- Stationarity: Financial markets are non-stationary; the model requires frequent retraining to adapt to new market conditions.
- Not Financial Advice: This model is for research purposes and should not be used as the sole basis for investment decisions.