Skip to content
Notifications
Clear all

TIL: The 'trending up' arrow sometimes triggers on noise.

3 Posts
3 Users
0 Reactions
3 Views
(@maya7)
Active Member
Joined: 1 week ago
Posts: 9
Topic starter   [#3796]

Was analyzing a client's daily active user chart in Looker. The built-in "trending up" indicator fired, suggesting a significant upward trend.

Pulled the raw data to check. The last 7-day average was only 0.8% higher than the previous 28-day average. The spike was caused by two anomalous low days falling out of the 28-day window, not genuine growth.

**Key takeaway:** The algorithm seems sensitive to period boundary shifts. Always verify with the underlying numbers.

Example SQL to calculate both averages and avoid the blind spot:

```sql
WITH daily_counts AS (
SELECT
date,
COUNT(DISTINCT user_id) AS dau
FROM events
GROUP BY 1
)
SELECT
AVG(CASE WHEN date >= CURRENT_DATE - 7 THEN dau END) AS avg_last_7d,
AVG(CASE WHEN date BETWEEN CURRENT_DATE - 35 AND CURRENT_DATE - 8 THEN dau END) AS avg_prev_28d,
(avg_last_7d - avg_prev_28d) / avg_prev_28d AS pct_change
FROM daily_counts;
```

Visualizations can over-signal. Trust but verify.

— m7


Data is the best salesperson.


   
Quote
(@moderator_mel)
Trusted Member
Joined: 4 months ago
Posts: 29
 

Good catch on the Looker indicator. It's a classic example of why we push for digging into the methodology behind any automated signal.

Your SQL workaround is solid for a one-off check. For anyone building this into a recurring dashboard, I'd also suggest adding a buffer period or using a rolling comparison (like last 7 days vs. the 7 days before that) to smooth out these boundary issues a bit more.

Thanks for sharing the specific numbers - seeing that it was only a 0.8% move really drives the point home. Visualizations are great, but they often have a default sensitivity that assumes noise is signal.


No receipts, no trust.


   
ReplyQuote
(@hannahb)
Estimable Member
Joined: 1 week ago
Posts: 76
 

Oh, the rolling comparison idea makes a lot of sense. I'm still pretty new to setting up dashboards in my team's project tool, and I think I've been trusting the default alerts too much.

When you say "buffer period", do you mean like adding a few days gap between the two periods you're comparing? So you'd compare last 7 days to the 7 days before that, but maybe skip a day in between to avoid the boundary effect?

That 0.8% number really is tiny. It's easy to see a green arrow and get excited before checking.



   
ReplyQuote