Day 128: Monday’s Analogs

Posted by on 8 May 2010 | 2 comments

Day 128

When it comes to forecasting, meteorologists try to use as many tools as available.  Above is the output from the Map Analog Retrieval Systems (MARS) produced by the Storm Prediction Center.  The idea is to take a given forecast field and compare it to the same field from the past.  The top four analogs/matches are then examined for severe reports that occurred on the given analogs/matches.  From these reports a probability of severe weather is created based solely off historical data.  This is a form of pattern recognition…we’re just having computers find the pattern!

Above is the MARS output for Monday.  In the upper right the analog is based on the forecast 500 millibar , the lower left is based on the 850 millibar analogs, the lower right is based on precipitable analogs, and the upper left is a combination of the other three.  The 500 millibar and 850 millibar analogs have a similarity…8 May 2003.  For those who don’t remember, there was a F4 that went through Moore, Oklahoma…Now, a lot can change between now and then, but Monday certainly has the potential to be a big severe weather day.

Below is some information regarding MARS.

1) MARS is a prefect prog approach. That prog happens to be a special version of the GFS ensemble mean. This is a version that is being tested here at SPC. The ensemble includes weighted time-lags from prior operational GFS, and GFS ensemble runs. There are 19 GFS members and MARS uses the overall mean for the height fields and the precipitable water field at each forecast hour. The ensemble mean can “wash out” significant features such as sharp troughs and ridges, especially at medium to extended ranges as the spread between the individual ensemble members increases. Verification statistics for the SPC Medium Range Ensemble forecasts are not yet available.

2) MARS uses the NCEP Reanalysis data with a rather course grid resolution of 2.5 degrees. This is far less detail than what is needed to discern the mesoscale features that play a huge role in severe weather. However, synoptically evident cases may be captured *if* the model forecast is correct.

3) Gridded severe weather output is based on arbitrary probability values ranging from 5% to 45%, with 15% roughly corresponding to a SLGT risk in past SPC Outlooks. We have not verified what a 15% contour means when generated on a MARS forecast map. If you look at the historical outbreaks using MARS you’ll see values of 25% inside areas that verified at over 45%. Is this a “good” or a “bad” forecast? We hope to have statistical verification after the 2005 severe weather season. In the meantime, use the link below for a graphical verification for the latest MARS forecasts.

4) Only 3 components go into the MARS scheme; 500 height gradient, 850 height gradient, and precipitable water. Obviously, many more ingredients that go into making a severe weather episode!

5) The MARS gridded severe weather output is based on the SPC rough logs from 1979 to 2004. There are huge problems with these logs including report inflation during the last 15 years. There may be incredibly good pattern matches from the earlier part of this period that may have very few reports associated with them. A poor match on a later date may contain 50-100 small hail events. Remember, “garbage in-garbage out”.

6) Again, be careful! Look at the date matches on the web page and make sure (if you can) that the pattern being forecast is similar to the analog pattern found by MARS. The RMS numbers help a little but investigation continues on ways to better define whether a match is “good” or “bad”. Right now, only the top 4 “matches”, or analogs are used. It’s entirely possible that some days will have 8 great analogs and other days may have none. On those bad days it is possible that a lot of unreliable severe weather data is being gridded anyway. Also, the RMS values usually decrease in the extended period as the GFS ensemble grid changes to a lower resolution.

  • Mark Wheatley

    Thank you for posting this.

  • http://www.spc.noaa.gov Greg

    Point number 1 in the MARS information section is a little outdated. The underlying forecasts used to generate the analogs are the means directly from the NCEP Global Ensemble Forecast System (GEFS). The fields that are being used for searching purposes are the mean height gradient at 850 millibars and 500 millibars, and the ensemble mean precipitable water for the valid time.

    Insight may be gained by looking at the time of year that the matches occur relative to the time of year of the forecast, and whether all three of the “predictors” (e.g. 850 gradient, 500 gradient, and PW) are resulting in analogs where severe weather was observed. In other words, you may sometimes have a forecast that produces severe weather analogs based on the 500 millibar height pattern but no severe weather analogs are generated based on the forecast PW. That could suggest that not all the ingredients needed for a severe weather event are in place based on that particular forecast/analog combination. If, on the other hand, there are 500 millibar analogs showing up for days in December and January, based on a forecast valid for a day in May, you could say that the 500 millibar gradient being forecast for this day in May has a good match to past cool-season days. If at the same time there are springtime analogs for severe weather based on the forecast moisture (PW), one might surmise that the forecast height pattern at 500 millibars has wintertime characteristics but the moisture is more common to spring…a potentially volatile pattern.

    There appears to be only weak correlation between greater probabilities and greater concentration of verifying severe weather. However, the outer 5 percent contour generally does a very good job at corralling the observed severe weather *if* the underlying forecast is reasonably accurate.

    One mystery that has so far eluded explanation. MARS seems to do a better job indicating severe weather potential in the 3-5 day range than in the 1-2 day range based on statistical evaluation that I conducted on the forecasts a few years ago and continuing assessment.

    In conclusion, while this approach to machine-based pattern recognition may appear overly simplistic at first, there are multiple levels at which it can be utilized in the forecast process.