Some Notes On the Results About Scaling

Here, I scaled the train/val/test data using all the dataset rather than using only train dataset. Indeed, it will create a label leak problem to a small extend.

If you feel that you don't want this, please change the class (ChunkSequenceTimefeatureDataLoader) setting in function torch_timeseries/experiments/experiment.py._init_data_loader as:

scale_in_train=True

However, I have to emphasize that this will not change the conclusion of this paper, here is a result clip if you scale using only the training data:

ETTh1-P96-MSE

FAN: 0.37903

SAN: 0.38818

RevIN: 0.40132

About Data Split

We use a general split paradigm: train/val/test = 7/2/1 and use all the data entries. For ETTs datasets, previous popular split is using only 2 years data and split as 6/2/2. We have implemented this and provide a correspondent dataloader: ETTMLoader/ETTHLoader at dataloader.py. To run the code, below is an example:

bash ./scripts/FAN/ETTh1.sh

or

bash ./scripts/FAN/ETTm1.sh

We have already checked the implementation from data, code, results perspective, comparing to previous implementation e.g. Informer, PatchTST, TimesNet.

1 Frequency Adaptive Normalization For Non-stationary Time Series Forecasting

This is the offcial repository of paper "Frequency Adaptive Normalization For Non-stationary Time Series Forecasting"

2 Prepare Datasets

ETTh1, ETTh2, ETTm1, ETTm2, ExchangeRate, Weather, Electricity will be downloaded automatically

3 Run baseline&FAN

3.1 install requirements

1. to run our code, Please make sure your python version is >=3.8.
2. install all the requirements, we assume that you have installed torch according to your environment

pip install -r ./requirements.txt

3.2 run scripts

Please first source this init.sh script:

source ./init.sh 

or manually add this directory to your PATHONPATH environment variable

export PYTHONPATH=./

1. FAN

Please change the settings in the following for what you need.

# running FAN using FEDformer backbone with output length 96, 168, 336, 720 on dataset Electricity Traffic with input window 96, and hyperparameter k = 4
./scripts/run_fan_wandb.sh "FEDformer" "FAN"  "Electricity Traffic " "96 168 336 720"  "cuda:0" 96  "{freq_topk:4}"

2. baselines

Please change the settings in the following for what you need.

# running all baselines~(DLinear backbone) with output length 96, 168, 336, 720 on dataset ETTm1 ETTm2 with input window 96
./scripts/run.sh "DLinear" "No RevIN SAN DishTS" "ETTm1 ETTm2" "96 168 336 720"  "cuda:0" 96

3 use model checkpoints

1. download the results folder for the checkpoints.
2. run the corresponding command to evaluate the results, for example:

# run FAN with K=2 on ExchangeRate L=96 H=96
./scripts/run_fan.sh "DLinear" "FAN" "ExchangeRate " "96" "cuda:0" 96  "{freq_topk:2}"

4 Z-score ablation

run the following command to run the z-score ablation experiments:

# evaluate with z-score reverse scale
./scripts/run_fan_invtrans.sh "DLinear" "FAN"  "Traffic " "96"  "cuda:0" 96  "{freq_topk:30}"

# run with no z-score
./scripts/run_fan_wandb_scale.sh "DLinear" "No"  "Traffic " "96"  "cuda:0" 96  "{freq_topk:30}" "NoScaler"