N-BEATS: Neural basis expansion analysis for interpretable time series forecasting

Arxiv

N-BEATS: Neural basis expansion analysis for interpretable time series forecasting

Keras / Pytorch implementation

philipperemy/n-beats

Demo

philipperemy/n-beats

1. Abstract

1-1.   Improving accuracy by 3% over Smyl's winning solution in M4 competition.

1-2.   M4 competition summary paper: 
   https://www.sciencedirect.com/science/article/pii/S0169207019301128

2. N-BEATS

2-1.   Input window typically ranges from 2H to 7H (H: Forecast Period Horizon)

2-2.   Each stacked Block X tries to decrease the residual from the previous Block X-1 (This concept looks like Boosting).

2-3.   Forecasts are sum of the outputs from each Block X.

2-4.   An activation function of FC stack (4 layers) is ReLU.

2-5.   Operation in Block X (skyblue box in the figure above) is as equations bellow,

2-6.   2nd equation above is the origin of N-BEATS ( Neural Basis Expansion ... ) which transform coefficients to outputs.

2-7.   Expansion functions can either be chosen to be learnable or can be set to specific functional forms.

2-8.   Training on input windows of different length (2H, 3H, ..., 7H) for a total of six window lengths and ensemble of those can improve model performance.