HREF/SREF/NAM retirement:
I have seen some people provide some clarity on aspects of this, but since this is something at the SPC we have been working with for awhile, I wanted to address all aspects of this in one spot.
🧵
NOAA has been working toward a unified modeling system for awhile. The SPC has known about these changes and have been working for awhile to alleviate the impact to operations. SREF products have been shifted to GEFS, etc.
All of the new models have been rigorously tested in the Hazardous Weather Testbed every year and by SPC staff. This feedback has been a crucial piece of the model development cycle and has shaped the timeline and the plan.
For the longest time the FV3 cam and the RRFS as well as the REFS trailed considerably behind the current HREF and HRRR. You can see this in many of the HWT reports. However, in 2025, this gap no longer existed.
Naturally, SPC forecasters had been very concerned before this point, and we still are! However, these results are helping to alleviate some of those concerns. Keeping the HRRR running until RRFSv2 was huge for this.
The SPC REFS adds the HRRR and removes the 2 worst performing REFS members. This version of the REFS outperformed the HREF in 2025. This version of the REFS will be displayed on the SPC webpage in place of the HREF when it is retired.
This transition will still be challenging. The REFS is underdispersive (not surprising since it is single model core), and still creates very large storms. However, to a trained eye and using post-processing techniques, it should be as useful as the current HREF
In summary, under the constraints of what the new modeling suite requires, a lot of work has been done to alleviate the impact to operations.
In addition, the SPC has access to multiple experimental MPAS versions daily and in the HWT.
In addition, the SPC has access to multiple experimental MPAS versions daily and in the HWT.
Based on what I have seen thus far, I would expect considerable improvements over both the RRFS and HRRR when RRFSv2 is released.
None of this is perfect, and transition is hard, but just know there have been years of development going on behind the scenes to make this as minimally impactful to ops as possible in fact. SPC machine learning models are being retrained on the REFS as I type this.
It is natural to be concerned, especially when such valuable tools are being retired. However, I feel better knowing that the same SPC science support branch who built the HREF has been tackling this model transition for awhile and has built the best solutions possible.
Generated by Thread Navigator
Press ⌘ + S to quick-export