Badge source
Key findings: Combining the forecasts of the following models via the median leads to the best overall performance for the monthly data of the M3-Competition (Makridakis et al.): ARIMA, CES, Theta, and DeepAR.
Statistical forecasting models can be useful for capturing linear relationships in time series realizations; however, they often rely on certain distribution assumptions and struggle to capture more complex relationships in the time series data. Additionally, statistical models are known to be sensitive to outliers which produce forecasts that may misrepresent the data.
Deep learning models, on the other hand, can handle difficult pre-modeling problems such as multiple seasonal components, missing data, and non-linear trends; however they require more computational power, more time and resources put into tuning, as well as higher understanding of the data in order to execute them properly.
By using ensembles that combine the satisfactory forecasts of multiple models, the forecasting accuracy can be improved even further. This research aims to ensemble various high-performing statistical and deep learning models to potentially achieve greater forecasting ability.
- ARIMA (differencing all realizations)
- ARIMA (differencing via Cochrane–Orcutt estimation)
- Holt-Winter's Additive
- Holt-Winter's Multiplicative
- Theta
- CES (complex exponential smoothing)
- ES (exponential smoothing state space model)
- MLP
- DeepAR
- LSTM
- R (refer to here for the libraries used)
- Python (used to implement LSTM and DeepAR)
Top method
TODO (lesson learned)
TODO
TODO