Human review runs on a schedule — Monday deep-dives, month-end closes. The events worth reacting to run on their own clock: a geo breaks out on a Wednesday afternoon, a credential expires on a Saturday, a campaign's efficiency starts decaying the hour its budget doubles. A weekly review's average detection delay is three and a half days — not because anyone is lazy, but because that's the arithmetic of checking weekly.
Anomaly detection sounds exotic; the concept is domestic. A baseline is a learned normal band for each series — this app's ad revenue, that geo's net, this campaign's spend — aware of its weekday rhythm (weekends aren't Tuesdays), its trend, and its usual noise. An anomaly is a reading outside the band by enough to matter, for long enough to not be a blip.
The critical word is each. A portfolio with 54 apps, 42 countries, and 1,389 campaigns is tens of thousands of series. No human watches that; every series watches itself.
1. The mid-week breakout. ES revenue starts running +22% over baseline on a Wednesday. Flagged that afternoon, it's a budget-shift that rides the wave — the shift move from the geography playbook, executed while the window is open. Found the following Monday, half the wave is gone.
2. The slow leak. An eCPM declining 1–2% a day never trips a daily eyeball test — each day looks like noise. Against a trend-aware baseline, ten quiet days of decline is a screaming pattern. Slow leaks are where portfolios lose money invisibly, because they're specifically shaped to defeat human perception.
3. The silent failure. A revoked OAuth token, an expired credential, a source that stopped syncing — these don't produce bad numbers, they produce no numbers, and dashboards render absence as zero or as yesterday's stale value. Absence-of-data is the loudest anomaly there is, and the one humans are worst at noticing.
4. The correlated move. Spend up 40% and revenue up 6% is efficiency decay in progress — each number individually unremarkable, the pair alarming. Baselines that watch relationships, not just levels, catch the scale-too-fast mistake in days instead of at month-end.
Detection is the machine's job; judgment isn't. The flag says ES is +22% — it doesn't know a competitor just left the market, or that your new creative launched there, or that this is the third breakout this quarter and the pattern is the story. Verdicts, context, and strategy stay with you. The honest division of labor: machines watch, humans decide.
Anomaly detection has failure modes of its own. Baselines need history — the first weeks are calibration, not clairvoyance. And thresholds set too tight produce alert fatigue, which is worse than no alerts: a channel that cries wolf trains you to ignore it. Good systems tune for few, meaningful flags and accept missing the trivial.
That's the manual version, and it genuinely works at small scale. Its limit is arithmetic: bands per series times series per portfolio. Somewhere between five series and fifty thousand, the pencil hands over to the machine — which is the entire reason our AI watches every series so the 4-minute morning can stay four minutes. Either way, the principle holds: don't review on a schedule what you could watch continuously.
Connect your first platform in minutes and see what your portfolio actually earns — after spend, every day.