Anomaly alerts: when to alert and when it's just noise
Alert on falling revenue or rising returns without drowning in messages: choose the period, threshold, change or anomaly condition, wait time and recipients.
By The DataxLand team· · 6 min read
Automated alerts sound like the perfect solution: no need to open a dashboard, your phone buzzes when something happens. In the first week the owner reads every message. By the third week the Telegram group has thirty "revenue down 25%" messages — mostly on Monday mornings or after public holidays — and nobody reads them any more. When a real drop finally comes, the alert sinks among the meaningless ones.
A good alert has three properties: it is rare, it is right, and it comes with something to do when you receive it. This post covers the sources of noise, how to choose the period and the kind of condition, and a starter set of alerts for a multi-channel shop.
Why alerts turn into noise
| Source of noise | Example | How to avoid it |
|---|---|---|
| Numbers too small | A new channel with 5 orders a day has 3 today = "down 40%" | Check weekly instead of daily; set an absolute threshold |
| Day-of-week seasonality | Sunday is always lower than Friday | Compare like with like (week with week), not adjacent days |
| Data not in yet | Ad source not refreshed, fees still estimated | Check only after a data refresh; look at the data date |
| Known events | Sale days, holidays, planned stock-outs | Keep an event calendar; pause the alert or accept one message |
| Repeats | The same drop reported every hour | Set a wait time between two alerts |
| Wrong recipients | Return-rate alerts sent to the whole company | Send to the people who can act |
Three kinds of condition and when to use them
Crosses a threshold
"Weekly return rate above 8%." Use it when you know clearly which level is unacceptable — usually for rates (returns, cancellations, fees as a share of revenue) rather than totals (revenue, orders), because totals move with the season while rates are more stable. Take the threshold from your own numbers: look at the last few months and set it above the highest level of a normal month.
Changes vs previous period
"Weekly net revenue down more than 20% vs last week." Suits totals, when the period is long enough that day-of-week patterns don't distort it. Daily, it is almost guaranteed to be noisy for a small shop.
Anomaly
The system compares the current number with the metric's usual range and alerts when it drifts well away from it. Suits metrics that fluctuate with a rhythm, where a fixed threshold is either too tight or too loose. Tune the sensitivity: high sensitivity alerts earlier but sends more messages. Alert in both directions when a sudden rise is also worrying (for example ad spend shooting up).
Example: what a real drop looks like
The example uses the DataxLand sample data: a simulated scented-candle and essential-oil shop selling on Shopee, TikTok Shop and its website from 1 Aug to 29 Oct 2026. These are not a real customer's numbers.
In the four weeks from 21 Sep to 18 Oct, Shopee had 391 orders, an average of about 97.75 orders a week. The week of 19–25 Oct had only 41 orders.
Change = 41 ÷ 97.75 − 1 ≈ −58%An Anomaly alert on weekly Shopee orders fires exactly once, as soon as that week's data is complete: the drop is well outside the rhythm of the previous four weeks. Before reacting, the recipient should check, in order:
- Has the data arrived? Is the latest data date for the Shopee source yesterday? A source that stopped refreshing looks exactly like a drop.
- Is it cancellations? Valid orders falling because cancellations rose is a different problem from fewer orders being placed.
- Where is the drop? Break it down by product: is the whole shop down, or one or two SKUs (out of stock, hidden, a competitor cut prices)?
- Did something happen? A price change, paused ads, a fee change, the marketplace changing its ranking algorithm.
A useful alert doesn't just say "down 58%" — it leads the recipient to these four questions.
A starter set of alerts for a multi-channel shop
The thresholds below are starting points to tune, not industry standards. Replace them with numbers from your own history.
| Metric | Period | Condition | Send to | What to do |
|---|---|---|---|---|
| Net revenue | Week | Down more than 20% vs last week | Shop owner | Check the four questions above |
| Orders by channel | Week | Anomaly, medium sensitivity | Channel operator | Check visibility, stock, prices |
| Return rate | Week | Crosses your own threshold | Operations, customer care | Look at SKUs and return reasons |
| Fees as a share of revenue | Month | Crosses your own threshold | Accountant | Check the marketplace's fee notices |
| Ad spend | Day | Anomaly, both directions | Whoever runs the ads | Check budgets, new campaigns |
| Profit after fees | Month | Down more than 15% vs last month | Owner, accountant | Ask "why did profit drop" by channel |
Five or six alerts like these are enough for most shops. Add one only when there is a question you really want to be alerted about, and someone responsible for acting on it.
Rules for keeping alerts clean
- Every alert has an owner. If nobody is responsible, turn it off.
- Weekly checks are the default; daily only for costs (ads) or incidents that need a same-day response.
- Set a wait time between two alerts from the same rule, at least as long as the period.
- Review monthly: alerts that fire often without anyone acting get a looser threshold or get deleted.
- Read the message with its data date. An alert without a data date is easy to misread.
How DataxLand does it
On the Reports and alerts page, Alerts tab, you choose the metric, an optional filter (for example one channel), a daily, weekly or monthly period, and one of three conditions: Crosses a threshold, Changes vs previous period, or Anomaly with a Sensitivity setting and the Alert in both directions option. Wait between alerts (hours) defaults to 24. Test run shows the alert against current numbers before you turn it on. Alerts are checked after every data refresh, at most once per period, so they never fire on data that is halfway through a build. See Reports and alerts.
Messages go to Telegram, Slack or Microsoft Teams. In a group, recipients can ask Si a follow-up right under the alert to break the cause down by channel or product (Ask Si in Telegram). The Suggestions tab proposes common alerts. See also alerts and reports on the feature page and the retail operations dashboard.
FAQ
Should alerts be daily or weekly?
Weekly for revenue, orders and return rate at small and mid-sized shops. Daily for costs you need to stop quickly, such as unusual ad spend. Only very large shops with thousands of orders a day have numbers stable enough for daily revenue alerts.
Can anomaly alerts replace thresholds I set myself?
Not entirely. An anomaly catches a change against the usual rhythm; a threshold expresses a level you won't accept, even once it has become "normal". A return rate that creeps up a little every week won't look anomalous, but it will still cross your threshold.
Should I turn alerts off on sale days?
No need to turn them all off. Weekly alerts fire only once; just warn recipients that numbers will be off during sale week. Keep the unusual ad spend alert on, though — sale days are when budgets most easily overrun.
How many alerts is too many?
When recipients stop reading them all. One check: count last month's messages and how many led to a specific action. If most led to nothing, loosen thresholds or merge alerts.
Read next: Automated Telegram reports · 5 sales data errors to check · P&L dashboard: 8 metrics to track