Can forecasters skillfully predict tornado activity levels 2-4 weeks in advance?
That is the question I am actively engaged in at the moment. This post is the result of several questions and comments about my latest BLOG post found here:
I will not focus discussion here on the 'secret sauce' going into the forecasts, so please do not ask.
Instead, I'd like to show the framework I'm currently working with and present verification for previous forecasts. I welcome feedback and discussion on the product presented below.
Introducing Extended Range Tornado Activity Forecasts (ERTAF)!
Here is a table of forecasts I have made so far this year (latest forecast issued yesterday 4/12 at 21:20 UTC):
FCSTDATE - Forecast date of ERTAF issuance. Forecasts are issued by 2300 UTC.
FCST2START - Start date of the week 2 forecast. Valid time begins at 0000 UTC.
FCST2END - End date of the week 2 forecast. Valid time ends at 23:59 UTC.
FCST3START - Start date of the week 3 forecast. Valid time begins at 0000 UTC.
FCST3END - End date of the week 3 forecast. Valid time ends at 23:59 UTC.
CLIMO2 - Average (1985-2014) number of tornadoes for the week 2 valid period.
CLIMO3 - Average (1985-2014) number of tornadoes for the week 3 valid period.
ACT2 - Actual number of preliminary tornado reports for the week 2 valid period.
ACT3 - Actual number of preliminary tornado reports for the week 3 valid period.
PN2 - Percent of normal tornado activity for the week 2 verification (ACT2/CLIMO2).
PN3 - Percent of normal tornado activity for the week 3 verification (ACT3/CLIMO3).
ACT2C - Categorical level of actual week 2 tornado activity based on Forecast Key.
ACT3C - Categorical level of actual week 3 tornado activity based on Forecast Key.
FCST2 - My categorical level forecast for the week 2 valid period issued on FCSTDATE.
FCST3 - My categorical level forecast for the week 3 valid period issued on FCSTDATE.
VERF2 -Verification of FCST2 based on the Verification Key.
VERF3 -Verification of FCST3 based on the Verification Key.
- Counts of all tornadoes are likely not the best metric. Considering the use of (E)F1+ counts.
- Environments would be better, but haven't thought much about a verification framework.
- Tornado days (any day with a tornado) may help alleviate tornado count problems and be the best metric to use.
- Forecasts beyond FCST3 may be possible.
- 1985-2014 at least gives a 30-yr climatological record. Different periods could be used, but the 30-yr mark is neccessary given the weekly periods I am examining.
- Better verification forthcoming. This is quick and easy for now.
I have no idea how skillful these forecasts will be, but it's time to start pushing the envelope. This is simple, but seems like a logical place to start.
Caveat emptor: These forecasts are solely my own and do not represent an official forecast from any corporation or government agency. Please use caution when interpreting the results and respect their origin.