ILLUSTRATIVE · NASA / SVS · NOT ADITYA-L1 DATA
Model cardspersistence
Model card · reference point
Persistence
Predicts that the next label equals the previous one.
What is this?
A model card for Persistence, one of eight detectors evaluated on the frozen Aditya-L1 flare benchmark.
Why should I trust it?
Every figure below is read at build time from the committed benchmark artifact; none is transcribed. The judgements are traceable to the written evaluation protocol and the dataset limitations record.
Where can I verify it?
- The validation evidenceComputed from 192,541 held-out predictions
- Dataset limitationsL-1 … L-10
- Evidence traceabilityclaim → artifact → commit
- The raw payloadfindings/models.json
Section 1
Purpose and intended use
Why it is in the benchmark. The most important reference point on the board. It scores high not because it forecasts but because the label is autocorrelated — flares last longer than a minute. Any forecasting claim that does not beat persistence is a claim about autocorrelation.
Intended use. Mandatory baseline for every forecasting result, per L-2. Reporting a forecast AUC without it would make the number uninterpretable.
Out of scope. Not a detector and not a forecaster.
Section 2
Inputs, outputs, training
Inputs. The previous minute's label.
Outputs. A copy of the previous label.
Training protocol. None.
Operating point. Decision threshold 0.5000, selected on the training period and then frozen.
Section 3
Evaluation
Both tasks, both reported.
M/X NOWCAST · 192,541 held-out minutes · base rate 1.24%
| Metric | Value | What it measures |
|---|---|---|
| ROC-AUC | 0.982 | Ranking quality across all thresholds |
| PR-AUC | 0.932 | The informative curve at a low base rate |
| Precision | 0.966 | Of the minutes flagged, how many were flares |
| Recall | 0.964 | Of the flare minutes, how many were flagged |
| F1 | 0.965 | Harmonic mean of the two above |
| MCC | 0.965 | Correlation, robust to imbalance |
| Brier | 0.001 | Probability accuracy; undefined for hard classifiers |
| Balanced accuracy | 0.982 | Mean of the two class recalls |
- Events detected
- 82 / 82
- Event recall 1.000 · 95% CI 1.000–1.000
- False alarm runs
- 0
- Of 85 alarm runs raised. A run, not a minute — the unit an operator actually responds to.
- ROC-AUC interval
- 0.978–0.986
- 95%, day-block bootstrap.
M/X 30-MIN PREDICTION · 192,541 held-out minutes · base rate 1.29%
| Metric | Value | What it measures |
|---|---|---|
| ROC-AUC | 0.983 | Ranking quality across all thresholds |
| PR-AUC | 0.935 | The informative curve at a low base rate |
| Precision | 0.967 | Of the minutes flagged, how many were flares |
| Recall | 0.966 | Of the flare minutes, how many were flagged |
| F1 | 0.967 | Harmonic mean of the two above |
| MCC | 0.966 | Correlation, robust to imbalance |
| Brier | 0.001 | Probability accuracy; undefined for hard classifiers |
| Balanced accuracy | 0.983 | Mean of the two class recalls |
- Events detected
- 82 / 82
- Event recall 1.000 · 95% CI 0.976–1.000
- False alarm runs
- 0
- Of 83 alarm runs raised. A run, not a minute — the unit an operator actually responds to.
- ROC-AUC interval
- 0.982–0.984
- 95%, day-block bootstrap.
Confusion matrix · M/X NOWCAST
True negative
190,066
False positive
82
False negative
85
True positive
2,308
Section 4
Strengths and failure modes
Strengths
- Quantifies exactly how much of any apparent forecast skill is persistence.
Failure modes
- Cannot anticipate an onset — it can only report one after it began.
Section 5
Limitations and ethical considerations
Ethical considerations. None — it is a ruler, not a detector.
Governing limitation clauses
Clause text is not restated here. It is published in full on the dataset card, from the frozen DATASET_LIMITATIONS_FOR_ML.md report, so there is exactly one copy of every caveat.