Use of Causal Inference to Improve Model Predictions

The conventional approach to improve model predictions is through model calibration, which is often subject to the curse of dimensionality as model complexity increases. Instead of relying on calibrating model parameters, causal inference is used to close the gap between model predictions and observations utilizing the information from modeled data otherwise not available through measurements. An information theoretic quantity called causally conditioned directed information is used to identify the modeled variables that can have causal influence on the observations. Then, the causal relationship between the selected causal variables and observations is represented using a highly effective machine learning model, eXtreme Gradient Boosting (XGBoost). Finally, the trained XGBoost model is used to make predictions of the target variable. We illustrate the use of causally conditioned direct information and XGBoost to improve the prediction of daily edge-of-field (EOF) runoff using the modeled variables from the National Oceanic and Atmospheric Administration's National Water Model (NWM), which as of v2.1 is solely calibrated to streamflow observations, not surface variables such as EOF runoff. Overall, the results helped us gain insights into the capability of NWM to simulate EOF runoff through the identified causal variables, and demonstrated that the predictions of EOF runoff over all domains of interest are largely improved.