Figure 2
From: Time series causal relationships discovery through feature importance and ensemble models
Pipeline used to perform causal discovery. Firstly, a model is built considering only past-lagged values of the target itself, which defines the forecasting baseline (I). The second step is to build a model considering information from both the target and a potential driver that may causally affect the target (II). If there is an improvement in the forecasting, the driver is kept as a predictor (added to the list of discovered drivers, and the baseline is updated). When a new potential driver is tested, information from the already tested drivers that causally affect the target (discovered drivers) is also included.
