ML-pipeline

ML-pipeline are the common tasks necessary to train models using machine learning, and its summarized in figure bellow.

The ML-pipeline usually includes data processing aspects and sometimes also the deployment of the final model, but here ML-pipeline deals with building the model to forecasting. We reserved one chapter exclusively for data processing due to its importance in COVID-19 forecasting.

Methods

That are many methods of machine learning to find models to forecast time-series. These methods can be categorized in three simple categories: Baseline, Autoregression, Epidemiological and Machine Learning. Here the Machine Learning category encompass: Linear Machine learning, Nonlinear machine learning, ensemble machine learning, deep learning and many more variants.

Methods for time-series forecasting

Baseline	Autoregressive	Epidemiological	Machine Learning
Naive	A	SIR	Vannila LSTM
Average	AR	SIRD	Stacked LSTM
Exponential Smoothing	ARMA	tSIR	Bi-directional LSTM
Gomperz growth	ARIMA	SEIRD	Convolutional LSTM
Logistic growth	VARIMA		ANN
Exponential growth	MARS		Random Forest
von Bertalanffy growth	ANFIS		Autoencoder
Probability distributions	ANFIs		Support Vector regression
			Time Delay neural networks (TDNNs)
			Extreme Learning machine
			Prophet forecasting model

Training

These methods get some data as input and tune its parameters to reduce the error of the expected output and the model output. This is the training part. All methods produce a similar result, i.e, a model that can be used to make predictions about future outcomes.

Evaluation

Metrics

Ensemble

Combine forecasts from many methods to improve forecast performance.