Skip to content

ILLUSTRATIVE · NASA / SVS · NOT ADITYA-L1 DATA

Model cardslogistic

Model card · candidate detector

Logistic regression

A linear model over the fourteen features, squashed to a probability.

What is this?

A model card for Logistic regression, 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?

Section 1

Purpose and intended use

Why it is in the benchmark. The lowest-capacity learned model, included to separate 'learning helps' from 'flexibility helps'. If a linear combination matches an ensemble, the extra capacity was buying nothing.

Intended use. Understanding which features carry the signal, via coefficients that can be read directly.

Out of scope. Its coefficients are correlational, not causal, and the features are strongly collinear — several rolling means measure nearly the same thing. A negative coefficient is not evidence that the quantity suppresses flares.


Section 2

Inputs, outputs, training

Inputs. Fourteen features per minute, all derived from the SoLEXS 1-minute count rate in T1: the log rate, rolling means over 5/15/30/60 minutes, a 15-minute rolling maximum, rolling standard deviations, a background-excess term, two rise rates, and three data-quality terms (GTI fraction, seconds present, partial-minute flag).

Outputs. A probability per minute.

Training protocol. Chronological split with a held-out test period beginning at the frozen test start date; no shuffling, so no future minute can inform a past one. The protocol — split, seed, metric set and bootstrap scheme — was fixed in EVALUATION_PROTOCOL.md before any model was fitted.

Operating point. Decision threshold 0.9524, 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%

MetricValueWhat it measures
ROC-AUC0.964Ranking quality across all thresholds
PR-AUC0.556The informative curve at a low base rate
Precision0.540Of the minutes flagged, how many were flares
Recall0.537Of the flare minutes, how many were flagged
F10.539Harmonic mean of the two above
MCC0.533Correlation, robust to imbalance
Brier0.029Probability accuracy; undefined for hard classifiers
Balanced accuracy0.766Mean of the two class recalls
Events detected
80 / 82
Event recall 0.976 · 95% CI 0.929–1.000
False alarm runs
228
Of 320 alarm runs raised. A run, not a minute — the unit an operator actually responds to.
ROC-AUC interval
0.953–0.974
95%, day-block bootstrap.

M/X 30-MIN PREDICTION · 192,541 held-out minutes · base rate 1.29%

MetricValueWhat it measures
ROC-AUC0.780Ranking quality across all thresholds
PR-AUC0.120The informative curve at a low base rate
Precision0.155Of the minutes flagged, how many were flares
Recall0.356Of the flare minutes, how many were flagged
F10.216Harmonic mean of the two above
MCC0.220Correlation, robust to imbalance
Brier0.060Probability accuracy; undefined for hard classifiers
Balanced accuracy0.665Mean of the two class recalls
Events detected
38 / 82
Event recall 0.463 · 95% CI 0.268–0.600
False alarm runs
221
Of 254 alarm runs raised. A run, not a minute — the unit an operator actually responds to.
ROC-AUC interval
0.674–0.850
95%, day-block bootstrap.

Confusion matrix · M/X NOWCAST

True negative

189,053

False positive

1,095

False negative

1,107

True positive

1,286


Section 4

Feature attribution

Which of the fourteen features the fit relied on.

  • log_rate+5.90
  • roll_std_15-4.70
  • roll_mean_15-3.24
  • gti_fraction+2.30
  • roll_max_15+2.20
  • roll_mean_5-2.02
  • rise_5-1.22
  • roll_mean_60-0.89
  • roll_std_60+0.86
  • roll_mean_30-0.45
  • rise_15+0.31
  • q_partial-0.21
  • n_seconds_present-0.16
  • bg_excess-0.15
Signed coefficients on standardised features. A negative value means the feature pushes the score down once the others are held fixed — which, given how strongly these features correlate, is a statement about the fit and not about the Sun.

Attribution is not explanation. These features are strongly collinear — four of them are rolling means of the same series — so credit assigned to one is credit withheld from another, and none of it constitutes a physical account of flare onset.


Section 5

Strengths and failure modes

Strengths

  • Coefficients are directly readable, including their sign.
  • Low capacity, which matters when the effective sample size is events rather than minutes.

Failure modes

  • Collinear features split credit unstably between themselves.
  • A linear decision surface cannot express 'high rate but flat and quiet', which is what separates a real rise from a noisy plateau.

Section 6

Limitations and ethical considerations

Ethical considerations. Readable coefficients invite over-interpretation. Presenting them as physical mechanism would be the misuse this card exists to pre-empt.

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.