How Supernova Decides Something Is Bursting
A rising count is easy to draw. This report walks through the test Asterism runs before it calls a rise a burst, and the guards around it. It ends with what a burst does not tell you.

The Question Supernova Answers
Every morning someone asks whether a technique, a malware family or an actor is suddenly everywhere. Asked honestly, that question needs a baseline to compare against and a way of counting that does not reward repetition.
Supernova answers a narrower question on purpose. Has this entity's share of the reporting Asterism collected grown, compared with its own recent past, by more than chance would explain? Share is the important word. If a large vendor starts a new blog and posts ten times a day, every entity's raw count goes up. Its share does not.
Count Stories, Not Reports
Ten outlets rewriting one vendor post are one observation. Asterism groups the reports about one activity into a Story, and Supernova counts Stories by default. For one entity and a 7-day window that gives four numbers:
k, the stories in the window that mention the entity;b, the stories that mention it in the baseline, the 90 days before the window, with the same focus and the same visibility settings;E_wandE_b, all the evidence-bearing stories in the window and in the baseline.
Undated items and marketing posts are left out of all four. A backfilled indicator repository stamped with today's date is not news, and counting it as news once produced exactly the kind of false surge this report is about.
The Test
If nothing changed, the entity's mentions should split between window and baseline in proportion to how much was published in each. Treat both counts as Poisson with the same underlying rate, put a flat Gamma prior on that rate, and the window count given the baseline count follows a negative binomial. Its tail turns out to be exactly a binomial tail: the conditional test Przyborowski and Wilenski described in 1940.1
share of all stories that fall in the window
w = E_w / (E_w + E_b)
what the baseline predicts for the window
expected = (b + 1) · E_w / E_b
the exact one-sided p-value
p = P( Bin(k + b, w) ≥ k )
a normal-equivalent score, for reading only
z = probit(1 − mid-p)With the sample numbers, E_w is 412 and E_b is 4,980, so the window's share w is 0.0764. T1656 appears in 9 window stories and 21 baseline stories, and the baseline predicts 1.82. That is 30 stories in total. If T1656 had kept its usual share, nine or more of them would land in the window with probability 0.00029.
The figure's mean, 2.29, is the binomial's: all 30 stories spread by the window's share w. The expected count above, 1.82, comes from the baseline alone, with one added to b by the Gamma prior.
The test is exact because most entities are rare. With counts this small, normal approximations are unreliable, and the exact tail costs almost nothing to compute.
Dividing by E_w and E_b means a source that is added or goes quiet moves every entity at once, so it cancels out of the comparison.
Testing Many Entities at Once
The board never tests one technique. It tests every technique seen in the window or the baseline, 148 in this example, and at that scale a few will look surprising by luck alone. The p-values are corrected for that before anything reaches the board.
First, Tarone's testability filter, in the false-discovery form Gilbert gave it, drops the hypotheses that could not reach significance even in the best case.23 An entity with three stories in total can never produce a p-value small enough to matter, so it should not dilute the correction for everything else. Here 118 of the 148 are testable.
Second, the Benjamini and Hochberg procedure runs over the testable set and turns each p-value into a q-value.4 The board lists an entity only when q is at most 0.05, it appears in at least 3 stories in the window, and at least 2 independent organizations reported it. Several feeds from one organization count as one witness.
| technique | k | b | expected | × | p | q | outcome |
|---|---|---|---|---|---|---|---|
| T1656 Impersonation | 9 | 21 | 1.82 | 4.9 | 0.00029 | 0.034 | listed |
| T1566.004 Spearphishing Voice | 6 | 14 | 1.24 | 4.8 | 0.0030 | 0.18 | q |
| T1556.006 Multi-Factor Authentication | 5 | 12 | 1.08 | 4.6 | 0.0074 | 0.29 | q |
| T1621 MFA Request Generation | 4 | 9 | 0.83 | 4.8 | 0.014 | 0.41 | q, 1 org |
| T1219 Remote Access Tools | 2 | 1 | 0.17 | 12 | 0.017 | none | untestable |
| T1078 Valid Accounts | 14 | 150 | 12.49 | 1.1 | 0.37 | 1.00 | steady |
| T1110.003 Password Spraying | 2 | 0 | 0.08 | 24 | 0.0058 | none | untestable |
Four help-desk techniques rose almost fivefold, and only one is listed, which is what the test is for. With six stories, a technique cannot be told apart from luck while 118 things are being watched, and the board would rather say less than say something it cannot back. A ratio of 24 on two stories, as for T1110.003, means nothing at all.
Finding the First Day
A significant window says the share grew. It does not say when. For that, Supernova runs Kleinberg's two-state burst model over the daily shares, with the usual parameters s = 2 and gamma = 1, and reports the first day of the latest burst that overlaps the window as the onset.5 For T1656 in the sample, that is 29 October. When the window is significant but no single burst is worth the model's cost of entering the burst state, the onset is left empty rather than guessed.
Guards Around the Test
Most of the engineering in Supernova goes into not comparing unlike things.
- Only sources collected through the whole baseline and window are counted, so a feed added last week cannot pass for a surge, and one that went quiet cannot pass for a decline.
- If collection stalled before or during the window, the board says it cannot tell. It does not read an empty window as everything fading.
- With fewer than 7 collected baseline days, rising and fading cannot tell either, and the board names the date from which they can.
- The NEW badge needs 30 collected days before it says anything, and knowledge-base entries such as ATT&CK techniques are never new.
When the data cannot support a comparison, the board shows "cannot tell" where the number would be, and says why.
A Burst Counts Coverage, Not Attacks
A listed entity is being written about more than usual, beyond what chance explains. It does not mean attacks went up, that the reporting is correct, or that the organizations involved saw the same thing. A significant trend says the share of coverage changed, not that activity did.
The model also treats stories as independent, and they are not quite: one vendor can publish a run of related posts in a week. The rule that at least two organizations must report the entity exists because of that, and it is the main guard against one prolific source producing a burst on its own.
Run It Yourself
The Supernova board on the Observatory, Asterism's home dashboard, runs this test on whatever you are focused on. The same test is an AQL pipe, so you can run it on any slice of the corpus:
published:>now-7d | trending type=technique mode=rising
entity observed expected × z q orgs onset
T1656 Impersonation 9 1.82 4.9 3.59 0.034 5 2026-10-29
-- 1 listed · 118 of 148 testable · unit: stories · fdr 0.05 --Swap mode=rising for fading, or for emerging to see entities named for the first time. Adding unit=reports counts reports instead of stories, which you should do only when you mean to.
Every constant used here is one of Asterism's defaults, and the board, the pipe and the API share them. The Asterism page shows a sample board and the platform's other instruments.
References
- Przyborowski, J. and Wilenski, H. (1940). Homogeneity of results in testing samples from Poisson series. Biometrika 31(3/4), 313-323. doi:10.1093/biomet/31.3-4.313 ↑
- Tarone, R. E. (1990). A modified Bonferroni method for discrete data. Biometrics 46(2), 515-522. doi:10.2307/2531456 ↑
- Gilbert, P. B. (2005). A modified false discovery rate multiple-comparisons procedure for discrete data, applied to human immunodeficiency virus genetics. Journal of the Royal Statistical Society, Series C 54(1), 143-158. doi:10.1111/j.1467-9876.2005.00475.x ↑
- Benjamini, Y. and Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society, Series B 57(1), 289-300. doi:10.1111/j.2517-6161.1995.tb02031.x ↑
- Kleinberg, J. (2002). Bursty and hierarchical structure in streams. Proceedings of the 8th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 91-101. doi:10.1145/775047.775061 ↑
Corrections
- correction The note under Table 1 said three help-desk techniques rose almost fivefold. The table shows four: T1656, T1566.004, T1556.006 and T1621.
Spot an error? Write to us and we'll fix it here, with a dated note.