I think I would look to create something that tested in some sense the limits of homogenization methods whilst still retaining some realism.
I would take a forced component run such as c20c and then look to add change points that in the net removed that trend. This would penalize any algorithm that tended to introduce adjustments with a preferential zero bias.
The breaks I would add would be a mix of step like and slope like and a large number would have changes in seasonality and timeseries variance associated.
I'd have limited metadata and what metadata there was would be poor quality.
I would assume that most breaks were small (sigma <1K, in some cases perhaps <<1K) and that they happened fairly frequently (once every 5 to ten years say on average).
A number of breaks would be quasi-contemperaneous over countries and these would have very similar characteristics to each other.
There is documented evidence that these issues all to some extent pervade the network (e.g. US network move from stevenson screen to automated sensors happened largely within 5 years over 70% of the network).
So, whilst at the outer bounds of plausibility it would not be an entirely implausible error structure.
3 comments:
Most ideas are good. Especially including breaks much smaller than 1K. Many previous studies only inserted breaks with detectable sizes. In reality the distribution of break sizes is about Gaussian and thus includes many small breaks. Including small breaks influences both detection and correction (also of the larger breaks).
However, if had only one world, I really would make sure that it is realistic. Inserting breaks every 5 to 10 years is more than double that found in real data. Homogeneous subperiods are on average 13 to 22 years (in industrialized countries). If the algorithms fail on such a high break rate, we still would not know whether they would work fine in reality. Given that we have multiple worlds in the surface temperatures initiative, we could spend one of them with a very high break frequency, where I would think that 10 years is already very high.
Inserting a trend from a standard model run, would allow the homogenizer to peak at the truth before the deadline. A stochastic trend, including decreasing trends, may thus be better.
I would like a world that explored geographically clustered inhomogeneities. I would do this by making some regional network wide changes to the global set of analog-stations. In some cases these would be simultaneous and in some cases they would cluster around a period of a year. They may affect the entire regional network or a percentage of it.
Post a Comment