GribStream Blog
NBM v5 parallel datasets now available
GribStream now exposes NOAA's experimental NBM v5 parallel datasets with hourly guidance through 48 hours, QM QPF/PoP, upgraded snow and wind products, and new fire-weather fields.
On March 24, 2026, we added NOAA's public NBM v5 parallel family to GribStream as a separate experimental dataset line rather than folding it into the operational NBM archive.
That distinction matters. The parallel feed is where NOAA is already publishing the upcoming NBM v5 field families, but it is still experimental, it can change, and upstream only keeps a short rolling archive. Keeping it under separate dataset codes lets you validate the new guidance against the current operational NBM products without breaking existing backtests or production queries.
What is included
- NBM Parallel (CONUS):
nbmpar - NBM Parallel (Alaska):
nbmparak - NBM Parallel (Hawaii):
nbmparhi - NBM Parallel (Puerto Rico):
nbmparpr - NBM Parallel (Guam):
nbmpargu - NBM Parallel (Oceanic):
nbmparoc - NBM Parallel QMD (CONUS):
nbmparqmd
All seven datasets are sourced from NOAA's public noaa-nbm-para-pds bucket on AWS Open Data.
Practical positioning
This feed should be treated more like RRFS pressure levels than an established long-history operational archive:
- it is experimental
- GribStream currently retains only the last 5 days
- NOAA can change field inventories, cadence, and product structure while the parallel line is still in evaluation
That makes it useful for side-by-side validation, QA, model transition planning, and early product work around NBM v5, but not yet a drop-in replacement for operational NBM, NBM Alaska, NBM Hawaii, NBM Puerto Rico, NBM Guam, NBM Oceanic, or NBM QMD.
Why NBM v5 is a real upgrade
The easiest change to notice is time resolution. In NOAA's July 15, 2025 v5.0 change notice, the agency highlights that most hourly guidance extends from 36 hours to 48 hours. If you use NBM as an hourly time series for load, road weather, dispatch, or short-range alert timing, that is a meaningful upgrade by itself.
The bigger story is methodological. NBM v5 is not just a few extra fields dropped into the existing blend:
- QPF and PoP change materially: CONUS precipitation guidance moves toward quantile-mapped weighting that leans harder on higher-resolution inputs, and NOAA adds the WPC MMEBC QPF as a CONUS input.
- Temperature and moisture products get a deeper refresh: temperature, dew point, relative humidity, and 12-hour max/min RH move away from older decaying-average logic toward quantile-mapped processing, which is what enables cleaner percentile and exceedance guidance.
- Wind guidance changes in a way users will actually feel: deterministic wind speed and gusts move to NOAA's percentile-picking approach instead of simply taking the quantile-mapped mean.
- The upstream blend gets sharper: higher-resolution ECMWF deterministic and ensemble inputs, higher-resolution Canadian guidance, and ECAIFS all enter the v5 picture for key fields.
That is why this release matters even if you never query JFWPRB or WETGLBT directly. The underlying blend is being re-tuned in ways that can move thresholds, percentiles, and event timing.
Winter, fire weather, and derived fields
NOAA's v5 package also adds several of the products that tend to matter most in real workflows:
- calibrated 24/48/72-hour snow exceedance guidance
- probability-matched-mean snow and ice products
- snow depth and snow-depth exceedance guidance
- joint fire weather probabilities for combined wind/RH thresholds
- apparent temperature and CONUS CAPE percentiles
On the public parallel feed now exposed on GribStream, you can already inspect several new or newly useful fields directly, including JFWPRB, PWAT, PWTHER, SNOD, and WETGLBT.
Why the parallel feed is worth using now
If you rely on NBM operationally, the main thing to validate is not just whether a new variable exists. It is whether your hour-by-hour triggers, snow thresholds, wind logic, percentile workflows, and backtests shift once the v5 methodology becomes the default.
The parallel feed is the cleanest place to do that comparison today because it lets you hold the dataset code constant on your side while comparing the v5 branch against the current operational NBM family.