This weekend’s storm is forecasted to drop as few as 2 or as many as 20 inches of snow on New York City depending on whether you use the American model or European. That’s a 1000% variance!  And you might start to wonder about the reliability of the predictions when two industry standard models vary so wildly, and more importantly, why are we talking about it beyond the fact that we’re about to get snowed in!

Although they may be different, each model has their strength – the “American Model” aka the Global Forecasting System  can predict 16 days in advance, whereas the “European model”, aka the EMWCF only predicts 10 days in advance, but does so with greater accuracy.

And there’s a value in getting disagreeing predictions. Both confirm it will snow a lot, they’re just uncertain of the exact location and exactly how much snow it’ll drop.  They vary primarily because they look at slightly different time horizons and structure their feature modeling differently thus giving different weights to different inputs to their models. Meteorologists can use those differences to give context to their weather reports, “This storm will drop a ton of snow, we just aren’t sure if the central path is going through New York City or over the water.”

So how does this apply to our work at Pickaxe?

When we work on new types of predictions for clients, we do what the best meteorologists do and use use an ensemble model approach. We try out multiple models, assess their different accuracies, and build a prediction process that takes the most accurate approaches from each model. And when the models disagree on a particular prediction, we can let clients know what our level of certainty is in a particular prediction. And then over time we can see how the contracting feature weightings and methodologies have fared in terms of accuracy.