Improving Forecasting Ability of GITM using Data-driven Model Refinement