If I am running a regression to forecast and predict future sales, do I need to worry about transforming my predictor variables (ie: taking the log) if I only care about prediction and don’t need to interpret coefficients?
Thanks.
Ask yourself why the linear transformation is not considered a heavy assumption, and how many relationships you’ve seen before that are very linear. Linearity is not the norm in my practice of statistics, and models don’t fit and don’t predict accurately when I give all continuous predictors the default linear transformation. Regression splines are your friend here. Inaccurate predictions generally come from three sources:
- having the wrong variables in the model
- nonlinear additive effects
- omitted interaction terms, linear or nonlinear
4 Likes