Accurate modeling of operational risk is important for financial institutions to prepare for potentially catastrophic losses. The loss distribution approach requires losses to be grouped into risk categories and loss frequency and loss severity distributions selected for each category. The annual operational loss distribution is estimated as the compound sum of losses over each risk category, and regulatory capital equal to the 0.999-quantile of this distribution is set aside. In practice, this approach can produce unstable regulatory capital year-to-year as the selected loss severity distribution family changes. We promote using truncation probability estimates and a consistent quantile scoring function on annual loss data as criteria for selecting severity distributions. Additionally, the Sinh-arcSinh distribution provides a flexible family for modeling loss severities. Finally, we investigate the effect of collecting and analyzing all loss severities.