Full Model Vs. LASSO Models
Full model vs. LASSO models
After creating and evaluating the full model, a back model and step model were created using backwards selection and stepwise selection, respectively. Additionally, there were two LASSO models created. Both of these LASSO models were created using a weight of 4 (TPR) to 1 (FPR). Full Model 2 also has these weights, so we will use this model to compare to the LASSO models.
The first LASSO model created was using “lambda 1SE.” This model was very minimal, only using 5 explanatory variables, compared to the 23 explanatory variables used by the full model. The Full Model 2 had a TPR of about 0.59 and a FPR of about 0.20, while the LASSO 1SE had a TPR of about 0.52 and a FPR of about 0.14 -- which were quite similar …show more content…
All models vary in their performance metrics and could be used in different organizational/business situations.
Cost Function Weights: Analysis
While working through the Credit Data and creating a variety of predictive models, there a few interesting insights. However, one of the most clear insights is the importance of the weights in the cost function. The TPR weight and FPR weight can be altered, but altering the TPR weight is the only one that makes sense in the context of the credit default problem. Altering the TPR weight can provide drastic changes to the performance of the model.
When the weights are set at (1:1), the model will respond by putting equal importance on the TPR and FPR. This is a standard set of weights that we will use. However, in the context of the credit default problem, this is an insensitive choice of weight. The TPR is very low, and could cause a lot of problems for a credit default evaluation organization. So, the TPR and FPR were analyzed as the weight on the TPR was increased from 1 to 10 -- giving us weights of (1:1), (2:1), (3:1), and so on to …show more content…
At the same time, the FPR for the model also increases. The most dramatic example of this increase can be seen between the Full Model 1 (1:1) and Full Model 3 (8:1). Full Model 3 is an interesting model because it provides us a very accurate TPR of 0.9906. However, Full Model 3 also provides us a FPR of 0.97. Full Model 3 is inaccurately predicting a credit default very frequently, which means that potential safe credit lenders could be denied opportunity to credit due to the aggressive weights placed by the model. This could cause a huge loss of revenue for business, as they will turn away potential customers due to fear of
After creating and evaluating the full model, a back model and step model were created using backwards selection and stepwise selection, respectively. Additionally, there were two LASSO models created. Both of these LASSO models were created using a weight of 4 (TPR) to 1 (FPR). Full Model 2 also has these weights, so we will use this model to compare to the LASSO models.
The first LASSO model created was using “lambda 1SE.” This model was very minimal, only using 5 explanatory variables, compared to the 23 explanatory variables used by the full model. The Full Model 2 had a TPR of about 0.59 and a FPR of about 0.20, while the LASSO 1SE had a TPR of about 0.52 and a FPR of about 0.14 -- which were quite similar …show more content…
All models vary in their performance metrics and could be used in different organizational/business situations.
Cost Function Weights: Analysis
While working through the Credit Data and creating a variety of predictive models, there a few interesting insights. However, one of the most clear insights is the importance of the weights in the cost function. The TPR weight and FPR weight can be altered, but altering the TPR weight is the only one that makes sense in the context of the credit default problem. Altering the TPR weight can provide drastic changes to the performance of the model.
When the weights are set at (1:1), the model will respond by putting equal importance on the TPR and FPR. This is a standard set of weights that we will use. However, in the context of the credit default problem, this is an insensitive choice of weight. The TPR is very low, and could cause a lot of problems for a credit default evaluation organization. So, the TPR and FPR were analyzed as the weight on the TPR was increased from 1 to 10 -- giving us weights of (1:1), (2:1), (3:1), and so on to …show more content…
At the same time, the FPR for the model also increases. The most dramatic example of this increase can be seen between the Full Model 1 (1:1) and Full Model 3 (8:1). Full Model 3 is an interesting model because it provides us a very accurate TPR of 0.9906. However, Full Model 3 also provides us a FPR of 0.97. Full Model 3 is inaccurately predicting a credit default very frequently, which means that potential safe credit lenders could be denied opportunity to credit due to the aggressive weights placed by the model. This could cause a huge loss of revenue for business, as they will turn away potential customers due to fear of