Skip to content

IKKIM00/multi-horizon-forecasting-comparison-between-TFT-and-DL-methods

Repository files navigation

General Description

Updated version for stock, pm2.5 datasets multi-horizon forecasting using Temporal Fusion Transformers(TFT).

You can find TFT paper through this paper link(https://arxiv.org/pdf/1912.09363.pdf).

Used stock, beijing pm2.5 and energy transformer(ET) dataset for multi-horizon forecasting.

Download Dataset

You can download each dataset through below URLs.

Stock Dataset

https://finance.yahoo.com/quote/CSV?p=CSV&.tsrc=fin-srch

Also you can use the data uploaded as data/stock_data.csv

Beijing pm2.5 dataset

You can download it through UCI Machine Learning Repository.

https://archive.ics.uci.edu/ml/datasets/Beijing+PM2.5+Data

Also you can use the data uploaded as data/beijing_data.csv

ET dataset

You can download the dataset through this github link: https://github.com/zhouhaoyi/ETDataset

How to use

For Deep Learning(DL) Methods

Through each ipynb files, you can choose model you want to use.

For TFT Methods

Through ipynb files that have tft in the file name, you can see the whole procedure.

Before running cells, make sure to clone original tft github page: https://github.com/google-research/google-research/tree/master/tft.