|
What is MyRELR?
MyRELR is a SAS macro that runs RELR (Reduced Error Logistic Regression), which is a completely automated and very general patented regression method that models all types of independent and dependent variables in Choice, Rating, Forecast, Survival Time, and Forecast models. RELR
regression coefficients have a tiny fraction of the error observed in
other regression algorithms. This lack
of multicollinearity and related overfitting error gives stable,
parsimonious, accurate and interpretable RELR models that avoid the significant
risk in Stepwise methods that Leo Breiman famously nicknamed as the "quiet scandal of statistics". RELR returns extremely accurate models with very small training samples in Big Data high dimension data with tens of thousands of variables/interactions. Because it is highly accurate even with extremely small observation samples, MyRELR can effectively deliver Big Data high dimension predictive modeling solutions through down sampling. Customers tell us that they only require about 5000 observations for extremely accurate RELR models that are built with tens of thousands of candidate variables/interactions and even can be run on a desktop computer in reasonable time.
The unique aspect of RELR is that error is automatically modeled and reduced in a non-arbitrary and non-trivial manner as a part of the maximum probability model build. The RELR error modeling is a major improvement over regularization methods like LASSO and Ridge Regression because it is much more accurate and has no manual tuning (see Case Studies page where there are comparisons including public studies by independent users). Thus, RELR does not require large observation samples for highly accurate and stable models even with extremely large numbers of variables and even with interaction and nonlinear effects. The MyRELR SAS macro automates all tasks in predictive modeling including imputation, variable reduction, variable selection, interaction and nonlinear effect building and the coding of nominal independent variables.
What are some of RELR's other advantages?
Imagine a stand-alone regression algorithm that has the benefits of Ensemble Modeling used in the Jeopardy and Netflix competitions in terms of giving very accurate models that have relatively very little error in regression coefficients and prediction, but without the weeks or months of laborious implementation time.
Imagine a regression algorithm that does not have arbitrary parameters and automatically gives the most probable parsimonious variable selection solution, so all modelers will generate this same most probable model.
Imagine an easy-to-understand regression algorithm that allows you to get an accurate model with a small percentage of the training sample observations that standard regression algorithms would require.
Imagine that the parsimonious selected model has causal plausibility because it is very accurate and stable and interpretable with correct regression coefficient signs and because roughly the same model would be generated by an independent training sample.
Imagine never having to worry about significant overfitting and multicollinearity problems because the regression algorithm does not impose limits on the number of variables and often can gets much more accurate when more variables are entered as candidate variables, even though the final parsimonious selection model may have fewer than 5-10 variables.
Imagine never having to worry about time consuming cross validations that are ambiguous and sample dependent because RELR models do not overfit.
Imagine usually getting a lift in classification accuracy compared to your current models with this lift often getting dramatically greater with high dimension data (an increase as much as 25 KS statistic points has been independently reported in a very high dimension problem involving 80,000 total variables), although RELR does not require a large sample size for such results.
Imagine that more complicated interaction and nonlinear variables are only selected in the final variable selection if they are stable and independent from simple linear variables, so you avoid the uninterpretable complexity issues in Stepwise Regression variable selection.
These imagined scenarios are what is typically seen in RELR modeling.
How can I get started?
The best way to get started is with free online training and a free 30 day software trial by using the form on the right or by calling us at 314-968-1875.
SAS is a trademark of SAS Institute. MyRELR and SkyRELR are trademarks of Rice Analytics. |
|
