1 of 10

Time Series Forecasting with GluonTS�an introductory guide

Levi Kaplan

Ming Luo

2 of 10

Background�

GluonTS

Time Series Data

Data

Probabilistic Forecasting

Goal

Time

Value

A sequence of data with time order

​

Predict distribution instead of a single value

Probabilistic Time Series Forecasting Graph

3 of 10

Project Overview�

Input

Output

Model

Required fields (columns) :

1. Start Field

2. Target Field

Pre-built deep learning models (Estimators):

SimpleFeedFoward (MLP)

DeepAR(RNN)

​

​

The distribution of all possible time series outcomes

Objective

Takes uncertainty into consideration by providing a range of all possible values

4 of 10

Project Models�

Naïve Seasonal

  • Repeats the previous ”season” of data in next prediction

5 of 10

Project Models�

Naïve Seasonal

Simple Feed-Forward

Network

  • Repeats the previous ”season” of data in next prediction
  • Simple MLP Network
  • Hyperparameter tuning:
    • Num. Hidden Layers
    • Learning Rate
    • Batch Normalization

​

6 of 10

Project Models�

Naïve Seasonal

Simple Feed-Forward

Network

DeepAR

  • Repeats the previous ”season” of data in next prediction
  • Simple MLP Network
  • Hyperparameter tuning:
    • Num. Hidden Layers
    • Learning Rate
    • Batch Normalization

​

  • Auto-Regressive Recurrent Network
  • Uses RNN network and training over similar time series
  • No feature engineering or historical data needed

7 of 10

Project Findings�

Naïve Seasonal

  • Naively performed very well
    • M4 dataset very consistent
  • Worse on less-consistent data
  • Performance:
    • MAPE: 0.186
    • RMSE: 1985

8 of 10

Project Findings�

Naïve Seasonal

Simple Feed-Forward Network

  • Naively performed very well
    • M4 dataset very consistent
  • Worse on less-consistent data
  • Performance:
    • MAPE: 0.186
  • Best hyperparameters:
    • More than one 100-node layers
    • No Batch normalization
    • Learning rate 0.0001
  • Performance 100 epochs:
    • MAPE: 0.198
    • RMSE: 3152

9 of 10

Project Findings�

Naïve Seasonal

Simple Feed-Forward Network

DeepAR

  • Naively performed very well
    • M4 dataset very consistent
  • Worse on less-consistent data
  • Performance:
    • MAPE: 0.186
    • RMSE: 1985
  • Best hyperparameters:
    • Two 100-node layers
    • No Batch normalization
    • Learning rate 1e-2
  • Performance 100 epochs:
    • MAPE: 0.185
    • RMSE: 3152

​

  • Great performance for 100 epochs
  • Performance 100 epochs:
    • MAPE: 0.140
    • RMSE: 3277
  • Best overall performance

​

10 of 10

Thank You!