This is an implementation of SCINet using tensorflow and a work in progress. I want to explore the possibility of using SCINet to predict cryptocurrency prices and how they compare to traditional approaches such as an ARIMA.
SCINet is a novel architecture for time series forecasting proposed in this paper. See original paper for link to datasets.
- See applications.testing.sinewave.py for usage examples
- Obtained similar results on the ETD dataset (ETDataset-main/ETT-small/ETTh1.csv) used in the orignal paper but only with a batch size of 16 instead of 4. The cause of the discrepancy is unclear - pending investigation.
- Scored poorly on crypto data (mse ~= 1.5, ase ~= 0.8 when data is relative difference). Learning curve suggests model is underfitting, which is expected as the data contains only a few basic features and has undergone minimal feature engineering. No hyperparamters tuning either. The score should serve as a baseline for future improvements.