Probabilistic Forecasts Based on “Reforecasts” Tom Hamill and Jeff Whitaker and
Improving probabilistic forecasts Better ensembles –More members –Improved initial conditions –Higher resolution –Improved forecast models Statistical corrections of the NWP forecasts (our main point: this can improve forecasts so much that it deserves more attention)
A tool for exploring calibration: the CDC “reforecast” data set Definition of “reforecast” : a data set of retrospective numerical forecasts using the same model to generate real- time forecasts. Model: T62L28 NCEP MRF (now “GFS”), circa 1998 ( for details). Initial states: NCEP-NCAR reanalysis plus 7 +/- bred modes (Toth and Kalnay 1993). Duration: 15-day runs every day at 00Z from to now. ( ). Data: Selected fields (winds, geo ht, temp on 5 press levels, and precip, t2m, u10m, v10m, pwat, prmsl, rh700, conv. heating). NCEP/NCAR reanalysis verifying fields included ( Web form to download at ). Experimental PQPF:
Application: tercile probability forecasts Climatological distribution split into 3 equally likely bins. These categories are often called Below/Near/Above Normal “terciles”. NCEP Climate Prediction Center (CPC) operational product
Dashed lines: tercile boundaries Red points: samples above upper tercile Blue points: samples below upper tercile Solid bars: probabilities by bin count Dotted line: a fitted model, TBD What can we do with a long data set of observed and forecast anomalies? With our reforecasts, we have 25+ years of data. Let’s use old data in a 31-day window around the date of interest to make statistical corrections.
Example: floods causing La Chonchita, CA landslide, 12 Jan 2005 week-2 forecast 6-10 day forecast
Comparison against NCEP / CPC forecasts at 155 stations, 100 days in winter MOS-based Week 2 forecasts using low-res T62 model more skillful than operational NCEP/CPC 6-10 day!
results
Other examples of calibration using reforecasts Example: Decile forecasts of 850 hPa temps over US
Analog high-resolution precipitation forecast technique (actually run with 10 to 75 analogs)
forecasts now downscaled to 5-km using “Mountain Mapper” technique.
Analog example: Day 4-6 heavy precipitation in California, 0000 UTC 29 December UTC 1 January 1997
Skill as function of location Notes: (1)Less skill where it’s dry (climatological forecasts better here, tougher to beat). (2) Regions where precipitation analyses are poor are less skillful (snowy regions, poor coverage by gages & Doppler)
Skill as f(time of year)
Comparison against NCEP medium-range T126 ensemble the improvement is a little bit of increased reliability, a lot of increased resolution.
Regional Reforecasts based on NARR and 32-km Eta? Leverage Mesinger et al.’s Eta regional reanalysis. Run small (~5 mbr) ensemble to 3 days? 8 days? for ~25 years. Continue to run Eta in real time. Develop range of statistical products based on Eta reforecasts. Preliminary estimate: computationally very expensive. 100 K for disk storage at CDC. Need advocacy of users to make this happen.
Conclusions Possible to achieve near-perfect reliability, good skill by calibrating forecasts with many years of old forecasts Great results with low-res model; even better results with higher-res. model? Want your feedback on important products Continued development depends on your advocacy.
Hacienda Heights, CA mudslides, 22 Feb 2005 (also rain on snow event for intermountain west) 6-10 day fcst
6-10 Day Week 2