Do confidence levels' ranges have a high correlation with prediction error I've created four linear regression models each with different variables. 
I looked at the error rate: (actual-prediction)/actual and also on the confidence levels (90%). I've noticed that there is no correlation between the error rate as I measured it and the range's size. I would expect that as the error rate is smaller so does the size of the range and the opposite so I would like to know if the phenomena of lack of correlation between the two metrics is valid. I attach a chart to Illustrate it below:

 A: I'm not 100% clear on what you're asking (what are your confidence intervals on?) but I think I understand well enough to answer. Assuming you mean you constructed confidence intervals for the linear parameters in your OLS, I think I can give an example as to why this doesn't need to be the case.
We construct (a very typical) type of confidence interval called "Wald-type" confidence intervals by looking at 
$T= \frac{(\beta-\hat{\beta})^{2}}{Var(\hat{\beta})}$ 
and turning it into something like 
$\hat{\beta}-sd(\hat{\beta})p_{1-\alpha/2}\leq \beta \leq  \hat{\beta}+sd(\hat{\beta})p_{1-\alpha/2}$ 
where p is the quantile point of whatever distribution T follows, right? Imagine a situation where the true model is $y_{i}= \beta_{0}+\beta_{1}x_{i}$ and all of the $y_{i}$'s are greater than 1000. I could very easily fit a model where $\beta_{0}=\beta_{1}=0$. In this case my confidence intervals would have width 0 and the error rate would be horrendous.
The width of your confidence intervals is governed by two things, the variance of your estimate for whatever parameter you're fitting (which is in turned governed by the distribution of your data) and the value of the quantile you're looking at (that is the "confidence level" you talked about). The higher the variance of your data and the larger "confidence level" you're looking for, the wider your CI's will be.
