Multi-Model Fusion for Robust Time-Series Forecasting

1
Multi-Model Fusion for Robust Time-Series
Forecasting
Weizhong Yan
Industrial Artificial Intelligence Lab GE Global
Research Center Niskayuna, NY 12309
2
Outline

• Problem Description
• Datasets
• Challenges and modeling strategies
• Our Approach
• The Results
• Final Remarks

3
Dataset characteristics
Time series with seasonality, trend, and outlier
Non-stationary
4
Challenges and modeling strategies
A large number of time series with different
features. Manual, ad-hoc modeling strategies are
not working
A model-building strategy that can automatically
identify features (i.e., trend, seasonality, etc)
of time series and arrives in a forecast model
with robust accurate performance for a large
number of time series
5
Our Approach(1)
- Preprocessing
automatically
Feature identification Feature treatment
Outliers
Trend
6
Our Approach(2)
- Modeling
Generalized Regression NN
7
Our Approach(3)
- Why GRNN?
Its a variation of nearest neighbor
approach Forecast for an input is a weighted
average of the outputs in the training examples.
The closer an input to the training example, the
larger the weight of its corresponding output.

• Advantages
• Its a universal approximator
• Its fast in training (one-pass learning)
• Its good for sparse data
• Disadvantages
• It requires large amount of online computation
• It almost does not have any extrapolation
  capability (forecast is bounded by min max of
  the observations)

8
Results(1)
9
Results(2)
10
Results(3)
11
Results(4)
12
Results(5)
13
Results(6)
14
Final remarks

• Developing a robust time series forecasting model
  is a challenging task.
• Developing an automatic model building process
  that can be reliably applied to a large number
  of time series with varying features is even more
  challenging.
• When the number of historical data points is
  small, fusion of multiple simple models seems to
  work better than a single complex model does

Future work

• Using more GRNNs
• Optimally determining the tunable parameter,
  spread, for GRNNs

15
Thank you