es-rnn

Overview

This is a work-in-progress of implementing ES-RNN by Slawek Smyl, winner of the M4 competition, and using it in a heuristic manner for forex-forecasting.

This repo contains:

• Holt-Winters implementation.
• GRU residuals learner.
• Multiplicative & Additive time-series-reconstruction.
• Forecasting of the forex dataset; training as well as metric results can be found in the outputs folder.

Differences between Slawek's version and this one

The original intention was to reproduce the complete ES-RNN algorithm, and its prediction results, on the entirety of the M4-competition dataset. I still have a long way to go. What I have not yet implemented, and I would certainly like to, are the following

• 3d Holt-Winters implementation (that is, multiple series processing in one step)
• multi-series-model with shared trend LSTM
• additive time-series-reconstruction
• autoregressive learner
• blender module to merge predictions from multiple series
• quantile loss to get prediction intervals
• Slawek's loss function that optimises two losses: quantile loss + regularization

$$L_q(y, \hat{y}) = q(y - \hat{y} )_{+} + (1- q) ( \hat{y} - y)_{+}$$

• compare performance on benchmark dataset
• reattempt the forex-forecasting

Replicating results

To replicate the results of the es-rnn on the forex dataset:

1. Clone this repository to your local machine.

   git clone [email protected]:mdarm/neural-networks-project.git
   

2. Navigate to the project directory.

   cd neural-networks-project/es-rnn/
   

3. Simply run the main.py script.

   python main.py
   

Dependencies

• torch
• numpy
• pandas
• matplotlib

References

1. A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting
2. The M4 Competition: Results, findings, conclusion and way forward
3. M4 Competition Data
4. Dilated Recurrent Neural Networks
5. Residual LSTM: Design of a Deep Recurrent Architecture for Distant Speech Recognition
6. A Dual-Stage Attention-Based recurrent neural network for time series prediction
7. Euro foreign exchange reference rates