Anomaly detection for Excel. Outliers caught before the board does.
Anomaly detection for Excel usually means "eyeball the chart and hope". DataHub Pro scans every numeric column on upload using rolling z-scores and surfaces points that are statistically out of distribution — with the date, the expected range, and how many standard deviations they sit at. Useful for variance reviews, audit prep, reconciliation work, and catching the data-quality issue before it lands in the board pack.
Updated 7 May 2026
Why anomaly detection isn't standard in Excel reporting
Most teams know they should be checking for outliers in their data — strange months, missed accruals, coding errors, fraud. Most teams don't actually do it because the workflow is painful.
The traditional approach
- Build a chart of the time series.
- Eyeball it for points that "look weird".
- Build a simple z-score column with =STDEV / =AVERAGE.
- Compute thresholds manually. Apply conditional formatting.
- Repeat for every column you care about.
For one column, doable. For a 30-column trial balance, untenable. So most teams skip it and discover the anomaly when someone notices in a meeting two weeks later.
What anomaly detection should be
Drop in the spreadsheet. Every numeric column is scanned. Anomalies are flagged with date, value, expected range, severity. Click any flag to see the source rows. Fold into your variance review or audit pack.
How DataHub Pro flags anomalies
Six layers around the core rolling-z-score detection. Each adds context that makes the alerts actionable.
Rolling z-score on every column
Default 12-period rolling mean + stdev. Each new point's z-score (deviations from rolling mean) is computed; |z| > 2 flagged as warning, |z| > 3 as critical. Tunable thresholds and window.
Auto-applied to all numeric columns
On upload the platform scans every numeric column for anomalies. No need to specify which to check; the dashboard surfaces the top flags across the whole file.
Severity ranking
Anomalies ranked by absolute z-score and value impact. The top 5 are surfaced first; full list available on click. Useful for prioritising which to investigate.
Missing-period detection
Beyond outlier values, the platform detects missing time periods (e.g. a quarter with no data). Useful for catching coding gaps or system-export failures.
Click-through to source rows
Each anomaly links to the underlying rows. See the transactions that drove the spike. Speeds up investigation from minutes to seconds.
Inclusion in Auto Report
Anomalies surface in the Auto Report DOCX/PPTX as a dedicated section with severity, value and explanation. Stops material outliers being missed in the executive summary.
Where anomaly detection earns its keep
- Variance review. The biggest variances vs budget are usually anomalies. The platform flags them in the same view, with the source rows one click away.
- Audit prep / Y/E. Top anomalies become the audit work programme. Document the explanation for each in the working papers.
- Reconciliation. Differences between systems often show as anomalies in one of the two. Catch them in the file load rather than the reconciliation step.
- Fraud detection. Statistical outliers (transactions far from the rolling average) are a signal — not proof, but a useful filter for human review.
- Data quality. Missing-period detection catches export failures, system gaps, missed cutoffs. Better caught at upload than at month-end.
None of these are exotic uses — they're the everyday work that takes hours of manual chart-staring and minutes of rolling z-score.
Who needs anomaly detection in spreadsheets
Finance, FP&A, controllers
Variance analysis, month-end review, audit prep. Anomalies in the trial balance catch coding errors before they reach the audit; anomalies in cost lines catch missed accruals; anomalies in revenue catch unusual one-off items.
Internal audit and compliance
Statistical sampling for review. Anomalies in transaction data surface candidates for deeper investigation. Documented in working papers with the rationale.
Operations and revenue analysts
KPIs that are out of distribution often point to a problem (or an opportunity). Daily/weekly KPI reviews with anomaly flags surface what to investigate.
What anomaly detection won't catch
Statistical outliers aren't the same as "wrong" numbers. A genuinely large month (a one-off contract, a windfall sale) flags as an anomaly even though the data is correct. The platform's job is to surface anomalies for human review, not to decide what's wrong. Expect ~5-10% of high-z-score flags to be legitimate large months that needed surfacing for the commentary.
FAQs
What threshold defines an anomaly?
Default: |z-score| > 2 for warning (≈5% of points in normal distribution), |z-score| > 3 for critical (≈0.3% of points). Tunable in the settings; some teams use ±2.5 to balance recall and noise.
Can I configure the rolling window?
Yes. Default 12 periods (one year of monthly data; one quarter of weekly data). Tune to match your business cycle. Shorter windows are more responsive to recent changes; longer windows are more stable.
Does it work on financial data with strong seasonality?
Yes — but for highly seasonal data (think retail with a 3× Q4 spike), a deseasonalised z-score is more useful than the raw rolling z-score. The platform applies seasonal decomposition where it detects clear seasonality and computes z-scores on the residuals.
How does it differ from Excel's conditional formatting?
Conditional formatting highlights cells based on absolute thresholds you specify. Anomaly detection computes statistical thresholds based on the rolling distribution, so a number that's anomalous now might not be anomalous in three months as the underlying mean shifts. The dynamic threshold means false positives are lower and the alerts are more useful.
Can I exclude known one-offs?
Yes. Mark a flagged anomaly as "explained" with a note (e.g. "one-off Christmas bonus pool") and it stops being surfaced. The audit log retains the original flag and the explanation.
Does it work on revenue, costs, headcount, etc.?
Anything numeric. The platform doesn't make domain assumptions — z-scores are dimensionless. Revenue spikes, cost dips, headcount jumps are all flagged the same way.
Will it produce too many alerts?
The default thresholds are calibrated to surface roughly 1-3 anomalies per 100 points — typically the right volume for a human reviewer to assess in a few minutes. Tunable if your file has more or less "normal noise".
Can it detect anomalies across multiple columns at once?
Yes — "multivariate anomalies" (e.g. revenue and customer count moving in opposite directions) are detected via correlation residuals. Default mode is per-column z-score; advanced mode opens the multivariate view.
How does anomaly detection appear in the Auto Report?
A dedicated section: Anomalies Flagged. Lists severity, date, value, expected range, and a link to source rows. Editable text per anomaly so you can add the explanation before sending the report.
Can I export anomalies as a CSV?
Yes — every anomaly with its date, column, value, z-score and severity. Useful for working papers or pasting into a tracking spreadsheet.
Catch the next outlier before the board does.
Drop in your most recent finance or ops file. Every numeric column scanned. Free tier — no credit card.
See also: · For accountants · For finance teams · For insight agencies · For small business · · Home