Prediction vs performance

The 2013 Spencers Creek peak snow depth is in at 185.9 cm on 26 August. Time to check on the prediction model.

My pre-season prediction from back in April was 170 ± 50 cm, so that’s a tick. Not quite so fast though. My method uses multiple linear regression with a range of parameters with links to our snow season:

Spencers Creek peak depth (cm) = 1438 – 0.61 x year – 24.8 x AAO + 0.86 x SOI – 13.2 x IOD – 10.4 x PDO – 0.17 x last year’s depth

There are six parameters there (definitions are here). Of those, three are basically known pre-season: calendar year, the 2-year PDO and last year’s depth. The other three have to be estimated (winter average AAO, SOI and IOD). Of those, numerical prediction models with some skill are available for SOI and IOD. The tough one is the winter AAO, which, whilst it displays considerable coherence at multi-month timescales, is really pretty much a guess in April.

My April picks for June-July-August AAO, SOI and IOD were +1.0, 0.0 and -0.2. The actual outcomes were -0.3(!!), +7.2(!) and -0.3. All three of those are improvements for snow depth, the first two major ones. Plugging them into the regression equation gives a post-season “prediction” (or hindcast) peak depth of 207 cm (also at ± 50 cm), which is a bit further from the outcome. The result looks like this: