Is it normal for logistic regression, to have predictors which have good Wald's Chi Sq, but still bad performance? I have trying to build a logistic model with some 10 variables. All of those variables have Wald Chi sq value<500. All are highly significant p-value<0.0001.
Event rate is about 3%. 
The concordance: 55.1%, Discon:28%, Tied: 17%. 
Is this behaviour normal? What can I do, to decrease the ties and increase concordance?
 A: This is very normal.  Think about ordinary least squares linear models on a large sample size where small effects can be detected but the overall $R^2$ is only 0.07.  An over-simplified but somewhat valid interpretation would be that you can't predict individual outcomes very well but you can describe tendencies.  For a binary logistic model, the $c$-index (concordance probability; AUROC) of 0.55 is indicating very mild discrimination ability, i.e., lots of overlap in predicted probabilities between observations with $Y=1$ and those with $Y=0$.
Assuming the number of candidate variables is 10, you may not need to build a model but rather just to specify one.  Be sure to not assume linearity.  You can often get more predictive ability using regression splines to relax linearity assumptions.  Then there's consideration of interactions ...
Be sure to look at the model's pseudo $R^2$ and especially at the model likelihood ratio $\chi^2$ statistic with degrees of freedom equal to the number of candidate variables.  If the LR $\chi^2$ statistic is not significant the whole exercise is on thin ice.
