Dust on Dirt

A few days ago we discussed how the GFS was an outlier of the many forecasts for the storm that will come through northern Utah tonight (see Some Thoughts on Using Long-Range Forecasts).  Specifically, it was producing a good deal more precipitation than most members of the Utah Snow Ensemble, which is derived from the US Global Ensemble Forecast System (GEFS) and ECMWF Ensemble (ENS).  

Sadly, the latest forecasts continue to produce light precipitation and just a frosting of snow for Alta-Collins.  As an example, the six members of the Rapid Refresh Forecast System (RRFS) ensemble produce 0.25" of water equivalent or less and 3" of snow or less for Alta-Collins.  

The latest HRRR and GFS are also below those thresholds.  The 0600 UTC GFS is down to 0.08" of water and about an inch of snow.  Dust on dirt if it verifies, although the snow might be somewhat wet and not cold smoke.  

Sometimes an outlier forecast like the one from the GFS verifies.  Low probability doesn't mean no probability.  It can be helpful to recognize there is the possibility of something happening and many of the products we have on weather.utah.edu are designed to illustrate the potential of outlier events such as high precipitation intensity periods.  But knowing the odds puts that potential into context.  

One question I was asked about that post about using long-range forecasts is shouldn't we weight the ensemble members?  For example, lean toward the ENS control and the ENS ensemble over the GEFS since the ECMWF model is "better." Being an academic, whenever I'm asked a question, I think of the goblin Griphook in Harry Potter and the Deathly Hallows: Part 2.  

The skill of any forecast model decreases with increasing lead time (model skill might increase in the first few hours of a model forecast due to spin up, but we'll ignore that here).  Eventually, even the best model in the world has no skill relative to climatology, meaning you're just as good forecasting with the long-term mean or probabilities as using the forecast model.  

As a result, the performance of a better modeling system over another modeling system is maximized at shorter lead times and decreases with time.  I have illustrated this conceptually below.  

How quickly model system skill degrades is dependent on the variable being forecast.  One can skillfully predict temperature or geopotential height farther into the future than precipitation.  In addition, forecast skill in northern Utah is lower than in many other regions of the US, so the rate of decline in forecast skill is higher here. 

The last comparison of GFS and ECMWF model forecasts for precipitation at mountain sites in the western United States that I am aware was done by my group (see Caron and Steenburgh 2020).  For the modeling systems available at that time, the 12-36 hour precipitation forecast produced for SNOTEL sites by the ECMWF was clearly superior to the GFS.  However, by 60-84 hour lead time, the gap was much smaller and not statistically significant (based on a 95% confidence interval).  So for the GFS and ECMWF available at that time, at about day 3, the overall skill of their precipitation forecasts at mountain sites in the western US became somewhat similar.  

So, how much you should weight one modeling system over another when looking at an ensemble likely decreases with time.  It might make sense to weight the ENS over the GEFS at short ranges, but it's unclear is such weighting makes sense at longer time ranges, at least for precipitation in the mountains of the western US.  Rather than basing the Utah Snow Ensemble entirely on the ENS, we included the GEFS because studies have show that doing so results in better probabilities.

The discussion above is informed by experience, but ultimately the only way to know is to do the math.  Once again, I leave this for you to do.