One of a series, also see editions: #1, #2, #4, #5
Updated…
OK, here’s how the new prediction model including southern sea surface temperatures goes:
The correlation coefficient rises to a healthy 44% (the standard error falls to 47 cm). The new model really only misses badly around Pinatubo1 (1991 & 1992) and the 2006 wipe-out.
Basis:
My previous model used these parameters:
- calendar year, to model the downtrend
- winter average2 Antarctic Oscillation3 (AAO)
- winter average Southern Oscillation Index4 (SOI)
- winter average Indian Ocean Dipole5 (IOD)
- 2-year Pacific Decadal Oscillation6 (PDO)
- last year’s depth, to model the negative autocorrelation (and, crudely, a 2-year cycle).
The peak depth correlations with SOI, PDO, IOD and AAO look like this:
For the new model, I’m adding the moderately strong correlation with southern sea surface temperature6 (SST), which looks like this:
I’m also dropping the correlation with last year’s depth, because there’s no known fundamental mechanism to explain it, so it’s probably just a chance artifact, and, if so, it cannot be predictive. As before, I get the prediction equation from multiple linear regression of the observed peak depths with the revised set of parameters over the record since 1958 (the IOD I use is not available 1954-57). The result is:
Spencers Creek peak depth (cm) = 1248 – 0.49 x year – 19.3 x AAO + 1.63 x SOI – 11.9 x IOD – 9.06 x PDO – 130.9 x SST
To make my 2014 prediction, all that remains is to estimate the parameters and plug them in. That’s for another post.
Notes:
1. Mt Pinatubo in the Philippines erupted in June 1991; the biggest volcanic eruption of the 20th century. It seems like that would be too late to affect our 1991 snow season, but the peak snow depth was very late that year (293 cm on 26 September). There was hardly any snow in June, and we barely made it to a metre by the end of August. I believe Pinatubo did affect our 1991 snow — in September, a couple of months after the eruption. There’s little doubt that it affected 1992; by then the short-lived cooling effect was global.
It’s interesting that if the Pinatubo years (1991 & 1992) are excluded from the regression, the correlation coefficient rises to 52% and the standard error falls to 44 cm. Perhaps I need a “volcano” parameter in my model.
2. “Winter average” is the average of the June, July and August monthly figures.
3. Antarctic Oscillation (AAO), also called “Southern Annular Mode” or SAM), is a measure of how tightly the circumpolar winds (“polar vortex” in one usage) blow around the pole. A loose pattern (negative AAO) leads to more polar storms reaching southern Australia, and more snow. The winter average AAO is used.
4. Southern Oscillation Index (SOI) is the difference between Tahiti and Darwin surface atmospheric pressures expressed as monthly standard deviations x10. SOI is an indicator of the El Niño Southern Oscillation (ENSO), an east-west quasicycle in equatorial Pacific Ocean surface temperature and wind patterns which correlates with precipitation across much of Australia, including with alpine snow. A positive SOI is associated with more (and some say wetter) Australian snow. The winter average is used.
5. Indian Ocean Dipole (IOD) is an ENSO-like variation in the smaller Indian Ocean, which correlates with precipitation across southern Australia, including with alpine snow. The winter average is used.
6. Pacific Decadal Oscillation (PDO) is a long-cycle, mostly north-south variation in the western Pacific Ocean, closely related (but not equivalent) to yet another claimed mode called the Inter-decadal Pacific Oscillation (IPO). Negative long-average PDO is weakly correlated with more snow. The 2-year average to August is used.
7. Sea surface temperature (SST) is that in the Great Australian Bight and northwest Tasman Sea, averaged over the box: latitude 30-37°S, longitude 115-160°E, and expressed as degrees Celsius anomalies from the 1951-1980 mean, detrended about 2014. The winter average is used. (I detrend the SSTs to make the regression equation parameters appear more sensible, and to make them more comparable with the previous equation. Detrending doesn’t alter the regression outcome.)
