The repository provides the code for the paper 'Autoregressive Convolutional Neural Networks for Asynchronous Time Series' (https://arxiv.org/abs/1703.04122) submitted to ICML 2017), as well as general code for running grid serach on keras models.

Files 'nnts/models/{CNN, LSTM, LR, SOCNN}.py' provide code for testing respective models, with the last one implementing the proposed Significance-Offset CNN.

Parameters for grid search can be specified in each of the above files.

The repository supports optimization of the above models on artifical multivariate noisy AR time series and household electricity conspumption dataset https://archive.ics.uci.edu/ml/datasets/Individual+household+electric+power+consumption The dataset has to be specified alongside the paremeters in each of the files listed above.

1.(a) Generate artificial datasets by running 'python generate_artifical.py' or (b) Change the dataset to 'household.csv' in the appropriate model file. 2. Run any of the model scripts 'python nnts/models/{CNN, LSTM, LR, SOCNN}.py'

Feel free to contact Mikolaj Binkowski ('mikbinkowski at gmail.com') with any questions and issues.