Edouard-Legoupil
commented
6 years ago Brainstorming
This assume that additional indicators - features engineering were create before
Depending on the type of target variable, a series of candidate algorithm will be selected.
Stepwise Regression - Forward selection and Backward elimination :
- Run bivariate regressions on all variables available
- Eliminate independent variables that are not statistically significant at 10% (90% Confidence Interval)
- Rank variables by đť‘…2, the Coefficient of Determination.
- Check for non-linear relationship between X and Y
- Run Preliminary Tests: Variance Inflation Factor, F-tests etc…
Diagnostic Tests on potential final Models:
- Multicollinearity (Variance Inflation Factor)
- Model Mis-specification (Ramsey RESET)
- Heteroskedasticity vs. Homoskedasticity (Breusch-Pagan; White)
- Decision whether OLS, WLS, GLS or FGLS is preferred
- F-tests (for fit; for linear restrictions)
- Outliers (Bonferroni; Cooks Distance; Leverage)
- Normality in Residuals (Jarque-Bera)
Like #6 , based on an additional column to be documented in the xlsform or indicator sheet, it shoudl be possible to add some predictive model capacity.
The target to predict could be defined and then algorithm will run to find out the best predictor and algorithm...