- Data about average monthly revenue from 2000 stores around whole country.
- Gini coeff. of reve around 20%, with 50% of observation around average, very thin tails of distribution
- Explanatory variables: geolocation of competitors, traffic drivers, purchasing power, population etc. in cachment area (most scatter plots cloud-shaped with small correlation)
- some outlier cases
I've tried glm, svr, regression trees with boosting. All models predictions are only a little bit better then average (RMSE in avg model = 40k, RMSE in regression = 35k, reve avg = 130k), with small errors in centre and huge overestimation in left tail and underestimation in right tail.
Any idea for new model type or prediction technique that can fit better in tails without big disturbance in the center ?