Worst-case analysis is used among financial regulators in the wake of the recent financial crisis to gauge the tail risk. We provide insight into worst-case analysis and provide guidance on how to estimate it. We derive the bias for the non-parametric heavy-tailed order statistics and contrast it with the semi-parametric extreme value theory (EVT) approach. We find that if the return distribution has a heavy tail, the non-parametric worst-case analysis, i.e. the minimum of the sample, is always downwards biased and hence is overly conservative. Relying on semi-parametric EVT reduces the bias considerably in the case of relatively heavy tails. But for the less-heavy tails this relationship is reversed. Estimates for a large sample of US stock returns indicate that this pattern in the bias is indeed present in financial data. With respect to risk management, this induces an overly conservative capital allocation if the worst case is estimated incorrectly.