“All Models Are Wrong, but Some Are Useful”

I have found myself repeating this famous Box (1980) quote a lot lately, so much so that one colleague said he was going to keep a tally (perhaps implying that it is getting annoying). Maybe explaining the three main reasons here will help shut me up. Building a new model, especially one used for policy purposes, takes considerable time, effort, and resources. In justifying such expenditures, one inevitably spends a lot of time denigrating previous models. For example, in pitching the third Uniform California Earthquake Rupture Forecast (UCERF3) (http://www.WGCEP.org/UCERF3), criticisms of the previous model included fault‐segmentation assumptions and the lack of multifault ruptures. In the context of including spatiotemporal clustering for operational earthquake forecasting (e.g., Jordan et al. , 2011), another criticism has been that previous candidate models not only ignore elastic rebound but also produce results that are antithetical to that theory. For instance, the short‐term earthquake probabilities model (Gerstenberger et al. , 2005), which provided California aftershock hazard maps at the U.S. Geological Survey web site between 2005 and 2010, implies that the time of highest likelihood for any rupture will be the moment after it occurs, even for a big one on the San Andreas fault. Furthermore, Monte Carlo simulations imply that excluding elastic rebound in such models also produces unrealistic triggering statistics (Field, 2012). > Given all models are wrong, what we really hope is that any new model is more useful than its predecessors and that the value added exceeds the total development costs. While UCERF3 includes solutions to these and other issues, it also embodies questionable assumptions and approximations of its own. For example, simplified rules are used to quantify which multifault ruptures are possible, including a 5 km distance threshold that allows faults separated by 4.999 km to rupture together, but not those separated by 5.001 km. Although …