❄️ Good News for Snow Forecasting: SSUM-Snow Fusion-F1 Shows a Positive Gain Over Archived NBM
A simple fusion of archived NBM guidance and a frozen structural snow signal improved snow-event discrimination in a reserved 2026 historical test
Can a structural snow signal add useful information to already strong numerical weather guidance?
That was the question behind the latest SSUM-Snow experiment.
And the result is encouraging.
In a reserved January-March 2026 historical test, SSUM-Snow Fusion-F1 improved both ROC AUC and PR AUC over the declared archived NBM comparator. The test covered 720 station-date cases across eight U.S. stations, including 238 snow events.
The improvement is not huge.
But it is positive, reproducible through the published project pathway, and supported by source-parity verification.
That makes it worth sharing.
🌨️ What is SSUM-Snow Fusion-F1?
The idea is deliberately simple.
SSUM-Snow already had a frozen structural predictor called SPF1.
Fusion-F1 asks a practical question:
Can SPF1 contribute some useful information when combined with NBM?
The frozen formula is:
F1 = 0.75 * R_NBM + 0.25 * R_SPF1
In simple terms:
- 75% of the combined signal comes from the ranked archived NBM snow-event score.
- 25% comes from the ranked frozen SSUM-Snow SPF1 signal.
There are no station-specific adaptive weights and no month-specific tuning in the frozen rule.
The idea is:
strong weather guidance + separately constructed structural information -> test whether the combination improves discrimination
🌟 The encouraging result
The reserved test used historical January-March 2026 data.
The experiment contained:
720 station-date cases
238 snow events
482 non-events
8 U.S. stations
The result:
+----------+--------------+-----------+-------------+
| Metric | Archived NBM | Fusion-F1 | Improvement |
+----------+--------------+-----------+-------------+
| ROC AUC | 0.911647 | 0.927168 | +0.015521 |
| PR AUC | 0.837182 | 0.875097 | +0.037915 |
+----------+--------------+-----------+-------------+Fusion-F1 also produced a positive ROC-AUC difference at 6 of the 8 stations.
And both frozen bootstrap intervals for the ROC-AUC improvement remained above zero:
temporal bootstrap 95% = [+0.002058, +0.025198]
station bootstrap 95% = [+0.006764, +0.024041]
The station bootstrap is based on only eight stations, so it should naturally be interpreted with caution.
Still, the overall picture is encouraging:
NBM was already strong, and the frozen SSUM-Snow signal added a modest additional discrimination gain in this reserved test.
🧪 Why the test design matters
This was not a test where the model was allowed to see the answer first.
The 2026 experiment used what we call a future-of-the-past design.
Imagine going back to a historical date.
For each target date D, SSUM-Snow is allowed to use only information that would have been available before the snow outcome for that date was known.
The sequence is:
past observations + archived NBM forecast
then:
frozen SPF1 + NBM predictor
then:
Fusion-F1 prediction
and only afterward:
reveal the already-finished historical snow truth
For the 2026 SPF1 reconstruction, the structural predictor used the declared D-2 through D-1 observation window. The target-date snow truth was excluded from predictor construction.
This gives us something useful:
the chronology of a real forecast experiment without having to wait months or years for the future to happen.
🔬 There was a negative result first
This is an important part of the story.
SSUM-Snow did not simply run tests until it found something positive and hide the failures.
Before Fusion-F1, the frozen SPF1 predictor was tested by itself against the archived NBM comparator in 2025.
It was clearly weaker.
+-----------+----------+----------+
| Predictor | ROC AUC | PR AUC |
+-----------+----------+----------+
| SPF1 | 0.760265 | 0.571326 |
| NBM joint | 0.910984 | 0.854979 |
+-----------+----------+----------+SPF1 versus the NBM joint comparator:
ROC AUC difference = -0.150719
Positive stations:
0 / 8
That negative result remains part of the SSUM-Snow evidence chain.
And that negative result changed the direction of the research.
Instead of asking:
Can SSUM-Snow beat NBM by itself?
the better question became:
Does SSUM-Snow contain some structural information that NBM does not fully capture?
Fusion-F1 was created to test exactly that.
💡 A simple change in perspective
The progression can be summarized very simply:
2025: standalone SPF1 < NBM
then:
75% NBM rank + 25% SPF1 rank
then:
2026 reserved test -> positive discrimination improvement
That is a much more interesting outcome than pretending the original standalone idea worked.
The structural signal appears useful as a complementary source of information alongside strong numerical guidance.
✅ The source data was checked too
A positive score means much less if the historical source reconstruction cannot be reproduced.
So SSUM-Snow also includes a source-parity audit.
The reconstructed historical predictors and truth records were compared against the frozen experiment receipts.
For the tested components:
2025 SPF1: 1200 / 1200, 0 mismatches
2025 NBM: 1200 / 1200, 0 mismatches
2025 truth: 1208 / 1208, 0 mismatches
2026 NBM: 720 / 720, 0 mismatches
2026 truth: 720 / 720, 0 mismatches
2026 prequential SPF1: 720 / 720, 0 mismatches
Overall:
SOURCE PARITY: PASS
This matters because the current result is not based only on a summary table.
The project preserves the frozen specification, evidence identities, source reconstruction path and verification software.
⚙️ The core formula is surprisingly small
Fusion-F1 itself is not complicated:
F1 = 0.75 * R_NBM + 0.25 * R_SPF1
The difficult part is everything around it:
- keeping predictor definitions frozen;
- preserving chronology;
- handling historical missing data consistently;
- keeping target truth out of predictor construction;
- preserving the rank-reference population;
- reconstructing upstream sources;
- and making the resulting evidence independently inspectable.
The reference implementation intentionally separates that small fusion rule from the larger historical reconstruction pipeline.
🧰 Verification is built into the repository
The current package verifier reports:
24/24 PASS
The reference implementation self-test reports:
3/3 PASS
The source-reproduction self-test reports:
10/10 PASS
And the repository now also has a passing GitHub Actions verification workflow.
The verification and reference programs use only the Python standard library.
So someone examining SSUM-Snow does not need a large Python dependency stack just to verify the frozen package.
⚠️ What does this result actually mean?
This is where careful wording matters.
The demonstrated result is:
SSUM-Snow Fusion-F1 showed a modest but consistently positive reserved-period improvement in snow-event discrimination over the declared archived NBM comparator.
That is encouraging.
But it does not establish that SSUM-Snow universally beats NBM.
It does not establish that SSUM-Snow beats modern weather forecasting generally.
It does not establish superiority for:
- snow depth;
- calibrated probabilities;
- operational forecasting;
- every location;
- every winter;
- every forecast horizon;
- or every possible NBM configuration.
And Fusion-F1 is not a standalone weather model.
It explicitly uses NBM as an input.
So the current claim is intentionally narrower:
the tested frozen structural signal added useful snow-event discrimination information to the declared archived NBM comparator during the reserved 2026 experiment.
🌨️ Why might this be interesting?
Modern numerical weather prediction already incorporates enormous amounts of atmospheric information.
That makes the positive result more interesting, not less.
The interesting question is:
Can a differently constructed structural signal identify additional information that remains useful even after strong numerical weather guidance is already present?
In this experiment, the answer was:
yes — modestly.
That opens a research direction around complementary structural information and measurable predictive improvement.
🔍 Inspect it, reproduce it and challenge it
The SSUM-Snow repository contains the current Fusion-F1 package, methodology, scientific status, claim boundaries, frozen specification, reference implementation, evidence files, source-parity result and source-reconstruction software.
🔗 Explore the SSUM-Snow Repository
Independent outside-party end-to-end replication remains:
OPEN_NOT_YET_CONFIRMED
Researchers, meteorologists, programmers and reproducibility enthusiasts are welcome to:
inspect it • reproduce it • question it • challenge it
That is exactly what the repository is there for.
🌐 Part of the Shunyaya Framework
SSUM-Snow is part of the wider Shunyaya Framework, which explores deterministic structure, reproducibility, structural signals and bounded scientific claims across several domains.
For SSUM-Snow, the current story can be reduced to four lines:
frozen structural snow signal
+ archived NBM information
+ chronologically reserved historical test
-> modest positive discrimination improvement
The earlier negative standalone result remains visible.
The later positive fusion result remains bounded.
And the evidence is available for examination.
✨ Closing Reflection
The most encouraging part of this result is that a frozen SSUM-Snow structural signal that performed poorly by itself appears to contain useful information when combined with an already strong NBM predictor.
That is a subtler — and potentially more interesting — result.
That sharpened the research question:
Can SSUM-Snow add predictive information beyond NBM?
And, more specifically:
Can that structural information improve prediction when fused with NBM?
In the reserved January-March 2026 historical experiment, the answer was modestly positive.
ROC AUC: 0.911647 -> 0.927168
PR AUC: 0.837182 -> 0.875097
6 of 8 stations positive
Both frozen ROC-AUC bootstrap intervals positive
Source parity: PASS
That is not the end of the research.
But it is a meaningful result worth putting in front of others.
❄️ NBM + Frozen Structural Signal -> A Small but Positive Gain
SSUM-Snow Fusion-F1 v1.0.1
F1 = 0.75 * R_NBM + 0.25 * R_SPF1
🔗 SSUM-Snow — GitHub Repository
Scope: Research • Reproducibility • Historical Evaluation
Outside replication remains open.
Comments
Post a Comment