From: Using passive Wi-Fi for community crowd sensing during the COVID-19 pandemic
VAR, VMA and VARMAX | VECM |
---|---|
1) Apply logarithmic transformation to the data, to guarantee that all forecasts are positive; | |
2) Apply differencing of 7 to remove seasonality \(z_{t} = y_{t} - y_{t} - 7\); | 2) Train VAR model to choose lag order with AIC; |
3) Apply differencing of 1 to remove trend (if needed) \(z_{t} = yt - y_{t} - 1\); | 3) Select cointegration rank, using the Johansen Cointegration test; |
4) Train model (60% of data) and choose lag order based on AIC | 4) Train VECM model with the selected lag order and cointegration rank using train set (60% of data); |
5) Test best model (lowest AIC) using the test set (40% of data); | 5) Test model using the test set (40% of data); |
6) Test the best model (lowest AIC) using the test set (20% of the data) |