Bayesian time series prediction

End to end implementation of paper Deep and Confident Prediction for Time Series at Uber in PyTorch. We use the Metro Interstate Traffic Volume multivariate time series dataset for training and eventually predicting traffic volume.

We make use of the implementation of variational dropout from keitakurita/Better_LSTM_PyTorch for the LSTM layers with dropout.

Workflow

Code is prototyped in the notebooks before transfer into cleaned up Python scripts for reuse.

• notebooks/01_dataset_creation.ipynb

  • Download dataset from UCI repository, preprocessing, cleaning and feature creation
  • → src/data.py

• notebooks/02_encoder_decoder.ipynb

  • Prototype the encoder-decoder pretraining part of the model
  • → models/encoder_decoder.py
  • → src/utils.py

• notebooks/03_encoder_decoder_dropout.ipynb

  • Prototype the encoder-decoder pretraining part of the model using LSTM cells with dropout
  • → models/encoder_decoder_dropout.py

• notebooks/04_pretraining_hyperparam.ipynb

  • Use Ax for guided hyperparameter search in the pretraining of encoder-decoder. We use GCE compute for GPU acceleration
  • → src/utils.py

• notebooks/05_pretraining_embedding.ipynb

  • Taking a look at the learned pretraining embedding
  • → src/utils.py

• notebooks/06_prediction_network.ipynb

  • Prototype the prediction network and run Ax for hyperparameter search
  • → src/utils.py
  • → src/data.py
  • → models/prediction.py

• notebooks/07_full_inference.ipynb

  • Implement the full inference including uncertainty calculation algorithms, run the test data through inference and plot results
  • → src/inference.py

• notebooks/08_evaluation.ipynb

  • Evaluate predictions on the test set by comparing with predictions from facebook/prophet
  • → src/evaluation.py

Results

In notebooks/08_evaluation.ipynb Prediction results on the test set are compared to those made with facebook/prophet.

Results on the classical time series prediction evaluation metrics are presented below:

Metric	Uber	Prophet
Mean absolute error	280.47	680.98
Root mean squared error	490.92	955.85
Mean absolute percentage error	0.13	0.41
Symmetric mean absolute percentage error	0.029	0.024

Time series predicitons using the full inference algorithm including uncertainty bounds:

Requirements

numpy
pandas
torch
tqdm
matplotlib
ax-platform
fb-prophet