r - Is cv.glmnet overfitting the the data by using the full lambda sequence? -

cv.glmnet has been used by most research papers and companies for similar functions such as cv.glmnet for Glmnet.cr When creating (a similar package that applies lasso to successive regression continuity ratio) I have come to cv.glmnet in this problem.

  cv.glmnet fits the first model: glmnet.object = glmnet (x, y, weight = weight, offset = offset, lambda = lambda, ...)  < / Pre> 
 The whole figure is created with the glmnet object goes as follows:  lambda  is removed from the full model fit  
  lambda = glmannet.object $ Lambda   
 Now they ensure that the number is greater than 3  
  if (nfolds <3) stop ("nfolds be larger than 3 ; Nfolds = 10 recommended ")   
 A list has been created to store cross-authenticated results  
  outlists = as.list (seq (nfolds))   
   i  seq (nfolds)) a  for the loop  to fit various data parts {which = foldid == i (Is.ma tricks (y)) y_sub = y [! Which,] and y_sub = y [! Which] if (is.offset) offset_sub = as.matrix (offset) [! Lambdas full data outlists [[i]] = glmnet (x [! Joe, drop = false], y_sub, lambda = lambda, offset = offset_sub, weight = weight [!] For which to use, and offset_sub = NULL # ], ...}}}   
 What happens if the data is correct to complete the data, then the whole data is made with lambdass. Can anyone tell me how could this probably not be more suitable for data? We want in cross-valuation that the model does not have any information about the left side of the data. But  cv.glmnet  cheats on !!!!!!!!!!   
 
 
 You are right that a cross-vanished measurement of the fit to take the "best" value of the tuning parameter Using an optimistic bias in that measurement is introduced when any format with a "best" value is a sample variant of the model, when out-of-sample performance is estimated, but  more fitting  Pre-defined values ​​Maijheshn (zero points) is caused by the fall-out-of-sample here.Sorry performance in comparison. It's unusual, in my experience - optimization is very constrained (in single parameter) compared to many other methods of convenience selection. In any case, the choice of tuning parameters, on an out-out set, or external cross-verification loop or bootstrapping is a good idea to validate the whole process. See.