We find that poor resolution of the tails of the distribution of integrands obtained inhibits convergence in Monte Carlo calculation of real time path integrals. We show that many methods previously tried to improve convergence neither resolve nor diminish the tails effectively. We find that large contributions to the integrand come from paths that have a large variance from the zero-path and/or that have few zero crossings. The results of crude dampings based on where such paths are poorly sampled suggest that exploring cancellations of paths characterized by path variance and zero crossings may be effective in improving convergence.