sdfsfdx / deepar

An Implementation of DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks

DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks

Description

This is an implementation of 1704.04110.

What this implementation does NOT contain

Two significant pieces are left out at this time, albeit trivial to implement.

1. The joint embedding learning for item categorization
2. The support for the Gaussian Distribution, suitable in forecasting of real valued timeseries.

• If you decide to implement the Gaussian Distribution, mind the rescaling of the distribution parameters. Refer to the paper.

Results

Since the paper does not provide quantitative results, we ran the tests with the carparts dataset on Amazon's Sagemaker. All the pre-processing & train/valid split was done exactly like stated in the paper.

SageMaker's output (single epoch)

[07/01/2018 14:22:34 INFO 139862447138624] #test_score (algo-1, wQuantileLoss[0.5]): 1.12679
[07/01/2018 14:22:34 INFO 139862447138624] #test_score (algo-1, mean_wQuantileLoss): 1.13427
[07/01/2018 14:22:34 INFO 139862447138624] #test_score (algo-1, wQuantileLoss[0.9]): 1.14175
[07/01/2018 14:22:34 INFO 139862447138624] #test_score (algo-1, RMSE): 1.07522369541

Notice, however, that these metrics are taken computing the ground truths in relation to the average of 50 samples

Our implementation is able to achieve RMSE 0.9215

Validation RMSE at each mini batch of the single epoch