with the skill
## Anomalies
| Date | Signups | Expected | Verdict |
|---|---|---|---|
| Thu 9/17 | **0** | ~320-345 (Thu range: 329-344) | Almost certainly a tracking or pipeline failure, or a signup outage |
| Fri 9/25 | **1,140** | ~300-315 (Fri range: 298-315) | About 3.6x normal. Likely a campaign or promo, bot or spam signups, or duplicated events |
Everything else is consistent. Weekdays run 298-344 (mean ~323) and weekends run 184-207 (mean ~197). There is a strong weekly pattern, and the data show a slight upward drift across the four weeks. With only 4 weeks, I wouldn't read much into that.
## Your outlier method matters
The weekend dip is real seasonality, not an outlier, so don't run outlier detection on the pooled 28 days.
- **Z-score (|z| > 3):** The 1,140 inflates the standard deviation (~182), so it flags 9/25 (z ≈ 4.6) but **misses the zero** (z ≈ -1.7).
- **IQR fences:** These catch both, but only because the anomalies are extreme. The method is blind to day of week. A bad Saturday of 120 would sail through.
Better approach: compare each day to its own day type (weekday vs. weekend, or the same weekday in other weeks).
## Before you drop them
Dropping outliers automatically isn't the right default. Find out what each one is first.
Show the rest (22 more lines)
- **The zero:** If the signups happened but weren't recorded, the data is wrong and should be fixed or backfilled from the source system. If there was a real outage, the day isn't representative of normal demand. In that case, exclude it from a "typical rate" or impute it.
- **The 1,140:** If it's bots or duplicates, remove it. If it's a real campaign, it's genuine signups and belongs in totals, but should be reported separately from the baseline rate.
## What it does to the number
| Approach | Mean daily signups |
|---|---|
| All 28 days as-is | 304.7 (median 316) |
| Drop both days (26 days) | 284.3 |
| Replace both with expected weekday value (~323) | ~287 |
The gap between mean and median on the raw data (305 vs. 316) shows how much the two anomalies distort the picture.
## Recommendations
1. **Match the method to the question.**
- If you want a baseline daily rate, exclude or impute the anomalies, and say so: "Excluded 2 of 28 days (7%): a 0-signup tracking gap on 9/17 and a 1,140 spike on 9/25 under investigation."
- If you want a monthly total, don't drop days. That understates the total. Use corrected data for the zero, and include the spike if it's real.
2. **Report weekday and weekend averages separately** (~323 and ~197), or a weighted figure. A single blended average depends on how many weekends fall in the window.
3. **Note the window.** These 28 days span September and October, so it isn't a calendar month.
If you can tell me what happened on 9/17 and 9/25 (outage logs, campaign calendar, traffic sources), I can help you decide how to treat each.