I'm going to perform a backtest on two VaR models on a very large dataset (+50.000 values). Normally, I would use the Christoffersen LR test but in my case, due to the very large number of observation, I'm afraid that the LR statistics would be 0/0, regardless of the empirical fraction of VaR failures (which is close to 2.8% at 1% VaR). So, I raised some doubts about my knowledge of the test... does it make sense to think this way or am I thinking wrong? Eventually, do you think is there another test that would fit my case? Thanks!