bitcoin_price_movement_transformer

Overview

This model is a Time-Series Transformer trained on historical Bitcoin (BTC) price data and volume metrics. It aims to predict the future price movement over a 24-hour horizon based on a rolling 7-day (168-hour) context window. It outputs a probabilistic distribution rather than a single point estimate to account for market volatility.

Model Architecture

The model implements an Encoder-Decoder Transformer specialized for sequential numerical data.

• Context Window: 168 hourly observations.
• Lags: Utilizes multi-scale lags to capture both immediate momentum and daily cyclical patterns.
• Output Distribution: Employs a Student's t-distribution head to better model the "fat tails" often seen in cryptocurrency market fluctuations.

Intended Use

• Market Analysis: Providing a baseline forecast for traders and financial analysts.
• Risk Assessment: Estimating the probability of significant price drops or spikes.
• Research: Studying the efficacy of attention mechanisms on high-volatility financial assets.

Limitations

• Exogenous Factors: The model cannot account for sudden news events, regulatory changes, or "black swan" events that occur outside of the historical price patterns.
• Financial Risk: This is an experimental tool. It should not be used as the sole basis for financial investment or trading.
• Stationarity: Crypto markets change rapidly; models trained on 2024 data may lose predictive power as market regimes shift.