weather modeling and forecasting in the context of weather derivative demand and supply. The vast
majority of extant weather forecasting literature has a structural “atmospheric science” feel, and although
such an approach is surely best for forecasting at very short horizons, as verified both by our own results
and those of numerous others, it is not obvious that it is best for the longer horizons relevant for weather
derivatives, such as twelve weeks or six months. Moreover, it is distributional forecasts, not point
forecasts, that are of maximal relevance in the derivatives context. Good distributional forecasting does not
necessarily require a structural model, but it does require accurate approximations to stochastic dynamics.
In this paper we took an arguably-naive nonstructural time-series approach to modeling and
forecasting daily average temperature in four U.S. cities, and we inquired systematically as to whether it
proves useful. The answer, perhaps surprisingly, was a qualified yes. Our point forecasts were always at
least as good as the persistence and climatological forecasts, but were still not as good as the
judgementally-adjusted NWP forecast produced by EarthSat until a horizon of eight days, after which all
point forecasts performed equally well. Crucially, we also documented and modeled the strong seasonality
in weather surprise volatility, and we assessed the adequacy of long-horizon distributional forecasts that
accounted for it, with mixed but encouraging results. We found, moreover, an interesting commonality in
the patterns of cross-city deviations from perfect conditional calibration, indicating possible dependence on
common latent components, perhaps due to El Nino or La Nina.
All told, we would assert that in the context of weather modeling as relevant for weather
derivatives, it appears that simple yet sophisticated time-series models and forecasts perform at least well
enough to suggest the desirability of additional exploration. When, in addition, one considers that time-
series models and methods are inexpensive, easily replicated, easily extended, beneficially intrinsically
stochastic, and capable of producing both point and density forecasts at a variety of horizons, we believe
that a strong case exists for their use in the context of modeling and forecasting as relevant for weather
derivatives.
We would also assert that our views are consistent with the mainstream consensus in atmospheric
-16-