DynaConF

This is the code accompying the paper "DynaConF: Dynamic Forecasting of Non-Stationary Time Series".

Dependencies

Install dependencies:

conda env create --file environment.yml

Synthetic Data Experiments

Generate synthetic data:

run/generate_synthetic.sh

Run univariate baselines:

run/synthetic_baselines.sh

Run multivariate baselines:

run/synthetic_baselines_mv.sh

Run our models NAR (StatiConF) and NNAR (DynaConF):

run/synthetic_our.sh

Generate the result tables

run/table_synthetic.sh

Results are stored in ./output/synthetic/.

Real-World Data (Set 1) Experiments

All the real-world datasets in Set 1 are from GluonTS.

Run our models NAR (StatiConF) and NNAR (DynaConF):

run/benchmark_our_static.sh

and then

run/benchmark_our_dynamic.sh

Generate the result tables

run/table_benchmark.sh

Results are stored in ./output/benchmark/.

Real-World Data (Set 2) Experiments

All the real-world datasets in Set 2 are publically available. Information of these datasets are in ./datasets/licenses.csv. We also include the processed datasets in ./datasets/, which can be used by copying the unzipped folder to ~/.mxnet/gluon-ts/datasets/.

Run our models NAR (StatiConF) and NNAR (DynaConF):

run/benchmark_new_our_static.sh

and then

run/benchmark_new_our_dynamic.sh

Run the baseslines

run/benchmark_new_baselines.sh

Generate the result tables

run/table_benchmark_new.sh

Results are stored in ./output/benchmark/.