This excerpt taken from the ALL 10-K filed Feb 23, 2006.
The nature and level of catastrophes in any period cannot be predicted
Although, along with others in the industry, we use models developed by third party vendors in assessing our personal lines homeowners exposure to catastrophe losses that assume various conditions and probability scenarios, such models do not necessarily accurately predict future losses or accurately measure losses currently incurred. Catastrophe models, which have been evolving since the early 1990s, use historical information about hurricanes and earthquakes and also utilize detailed information about our in-force business. While we use this information in connection with our pricing and risk management
activities, there are limitations with respect to their usefulness in predicting losses in any reporting period. Examples of these limitations are significant variations in estimates between models and modelers and material increases and decreases in model results due to changes and refinements of the underlying data elements and assumptions which lead to questionable predictive capability, and post event measurements that have not been well understood or proven to be sufficiently reliable. In addition, the models are not necessarily reflective of company or state-specific policy language, demand surge and loss adjustment expenses and the occurrence of mold losses, which are subject to wide variation by storm.