Notes to Tables and Figures
Table 1: We show each forecast’s root mean squared error, measured in degrees Fahrenheit.
Figure 1: Each panel displays a time-series plot of daily average temperature, 1996-2001.
Figure 2: Each panel displays a kernel density estimate of the unconditional distribution of daily average
temperature, 1960-2001. In each case, we employ the Epanechnikov kernel and select the bandwidth using
Silverman’s rule, h=O.9σN~02.
Figure 3: Each panel displays the residuals from an unobserved-components model,
η = Trendt + Sea9ona0t + ∑‰ P,Λ-,∙ + σ∕εr 1996-2001.
Figure 4: Each panel displays a kernel density estimate of the distribution of the residuals from our daily
average temperature model, Tt - Trendt - Seasonalt - pi∙T,f.i∙. In each case, we employ the
Epanechnikov kernel and select the bandwidth using Silverman’s rule, h=O.9σN~02.
Figure 5: Each panel displays sample autocorrelations of the squared residuals from our daily average
temperature model, ,Tt - Tregdt - Setasothtelt - jɪ p,.7ζ,,.j2, together with Bartlett’s approximate ninety-
five percent confidence intervals under the null hypothesis of white noise.
Figure 6: Each panel displays a time series of estimated conditional standard deviations (σf) of daily
average temperature, where ot = ɪ^i ^γc + γs^sin^2π<7-∣∣θj + Ctejl1 + βo^1, 1996-2001.
Figure 7: Each panel displays the ratio of a forecast’s RMSPE to that of a persistence forecast, for 1-day-
ahead through 11-day-ahead horizons. The solid line refers to the EarthSat forecast, and the dashed line
refers to the autoregressive forecast. The forecast evaluation period is 10/11/99 - 10/22/01.
Figure 8: Each panel displays the ratio of a forecast’s RMSPE to that of a climatological forecast, for 1-
day-ahead through 11-day-ahead horizons. The solid line refers to the EarthSat forecast, and the dashed
line refers to the autoregressive forecast. The forecast evaluation period is 10/11/99 - 10/22/01.
Figure 9: Each row displays a histogram for z and correlograms for four powers of z, the probability
integral transform of cumulative November-March HDDs, 1960-2000. Dashed lines indicate approximate
ninety-five percent confidence intervals in the iid U(0,1) case of correct conditional calibration.