This repository provides a step-by-step tutorial on the Temporal Fusion Transformer (TFT) for time series forecasting using PyTorch.
The project uses daily cryptocurrency data obtained from Yahoo Finance and includes a custom function designed to ensure full reproducibility of experimental results.
- DATA_PREPARATION.ipynb: data collection and preprocessing pipeline
- TFT_TUTORIAL_GITHUB.ipynb: model implementation and training
The dataset consists of daily financial time series, including Bitcoin prices and macro-financial variables such as gold, oil, equity indices, and exchange rates. Data are downloaded directly using the yfinance library.
A custom function is implemented to ensure reproducibility of results by controlling random seeds across all relevant libraries.
Install dependencies with:
pip install -r requirements.txt
The implementation is based on the original paper:
Lim, B., Arik, S. Ö., Loeff, N., & Pfister, T. (2021).
Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting.