Перейти к основному содержимому

Forecasts & anomalies

Path: HR Analytics Overview / Role Center → Charts

This family looks forward and for the unexpected. Forecasts project an indicator's future from its own history; anomalies flag periods that break the established pattern. There are two kinds of anomaly detection — a built-in statistical method and a machine-learning method — so you can pick whichever suits a given indicator.

How to read a forecast​

A forecast projects the coming periods of an existing indicator and is shown as a central estimate plus a confidence range (a lower and an upper bound). On a chart the estimate is a line and the bounds form a band around it.

Treat a forecast as guidance, not a commitment. The band shows the uncertainty; the wider it is, the less certain the projection. Forecasts of externally-driven measures (turnover, labour cost) are inherently less precise.

Forecasts are powered by the SMART ML Forecast service and require the source indicator to keep at least 24 periods of history (set its Lookback Periods to 24+). Each forecast names the statistical model it uses by default (ARIMA, ETS, STL for seasonality, or a combination) — this is configurable.

ForecastProjectsFromDefault model
Headcount ForecastTotal headcountWRK-HEADCOUNTARIMA
Turnover Rate ForecastRolling 12-month turnoverWRK-TURNOVER-12MARIMA
Payroll Fund ForecastPayroll fund (ФОП)PAY-FUNDETS+ARIMA
Payroll by Dept ForecastPayroll for the top 5 departmentsPAY-FUND (per unit)ETS+ARIMA
SSC ForecastEmployer SSC (ЄСВ)PAY-SSCETS+ARIMA
Total Labor Cost ForecastFund + SSC combinedPAY-FUND + PAY-SSCETS+ARIMA
Overtime ForecastOvertime hoursWT-OVERTIMEETS
Sick Leave ForecastSick-leave daysABS-DAYS-SERIES (SICK)STL
Vacation Days ForecastVacation daysABS-DAYS-SERIES (VACATION)STL
Absence Rate ForecastCompany absence rateABS-RATE-DEPTSTL
Labor Cost Budget Corridor12-month labor cost as a P10/P50/P90 probability corridor + three cumulative totalsPAY-FUND + PAY-SSCETS+ARIMA, percentiles 0.1,0.5,0.9
Replacement Demand ForecastExpected terminations next quarter (hiring-plan signal)WRK-HIRES-TERMS (TERM)Auto + intermittent detection
Vacation Liability ForecastVacation reserve, reported at the 90th percentile for accrual planningABS-LIABILITYETS+ARIMA, percentiles 0.5,0.9
Headcount by Dept ForecastHeadcount per organizational unit (global cross-series model — new/small departments are covered too)WRK-HEADCOUNT (per unit)Global (ARIMA fallback)

The last four ship disabled: enable them after the SMART ML Forecast connection is verified.

Advanced forecast options​

Four options extend any forecast where the setup card shows them. They require the current SMART ML Forecast service.

  • Forecast Quantiles — a comma-separated list of 2–3 percentiles in (0,1), for example 0.1,0.9 or 0.1,0.5,0.9. The published bounds then become those percentiles instead of the confidence band ("the value stays under X with ~90% probability"), and headlines switch to probabilistic wording. With three values the central estimate is the median.
  • Detect Intermittent — lets the service switch a sparse series (many zero months, e.g. per-department terminations) to a dedicated intermittent-demand model. Requires a single or empty (= Auto) Forecast Model — it cannot be combined with an ensemble.
  • Use Calendar Regressor — feeds the number of working days per month (from the official payroll calendar in Human Resources Setup) into the forecast as a known future factor, so a 19-working-day month is not mistaken for a downturn. Requires Forecast Model = ARIMA or AutoARIMA, and the payroll calendar must be generated for the whole history and the forecast horizon — a missing month produces a clear error naming the calendar and the month.
  • Use Global Model — for grouped forecasts (payroll or headcount by department): one cross-series model call covers all departments instead of one call per top-5 department, and departments with only a few months of history still get a forecast by borrowing strength from the others.

How to read an anomaly​

An anomaly is a period that deviates from what the history predicted. Each anomaly row shows the actual value, the expected value and — for ML anomalies — the deviation type (seasonal, level shift, change-point, or trend). Severity follows the size of the deviation:

  • 2–3 standard deviations → unfavorable;
  • above 3 standard deviations → critical;
  • a change against a perfectly flat history is treated as new behavior (ambiguous), not automatically critical.

Statistical anomalies (built-in)​

These use a z-score (mean ± a configurable number of standard deviations, Sigma Multiplier) over the trailing history — no external service required.

  • Payroll Anomalies (ANM-PAYROLL) — departments whose payroll fund deviates from their own history, employees whose period pay jumps versus their history (person-level, off by default; the history covers only whole months of their employment, so a recent hire is not compared with the months before hiring, and a period without any pay is left to Zero-pay Employees), and employees with earlier pay who received a payroll element they never had before (a new hire's first elements are not flagged). The anomaly count includes only what the instance shows: with person-level results off it counts the department findings. December / thirteenth-salary seasonality is handled by comparing the same calendar month of prior years once two years of history exist. Correction documents are excluded.
  • Absence Anomalies (ANM-ABSENCE) — a monthly company sick-leave spike against history (a possible seasonal outbreak), departments whose absence rate exceeds the company average by the configured sigma, and the count of long ongoing sick leaves (more than 20 calendar days).
  • Bradford Factor (ANM-BRADFORD) — see the responsible-use note below.

ML anomalies (SMART ML Forecast)​

These call the SMART ML Forecast anomaly service, which recognises seasonal, level-shift, change-point and trend deviations — richer than a plain z-score, and they name the deviation type. They are a machine-learning alternative/complement to the statistical anomalies over the same series; enable whichever suits the company. Each requires 24+ periods of history on its source indicator.

ML anomalyWatchesSource
Headcount Anomalies (ML)Workforce shifts (change-points)WRK-HEADCOUNT
Turnover Anomalies (ML)Turnover-rate deviationsWRK-TURNOVER-12M
Payroll Anomalies (ML)Payroll-fund deviationsPAY-FUND
Sick Leave Anomalies (ML)Sick-day deviationsABS-DAYS-SERIES (SICK)
Absence Rate Anomalies (ML)Absence-rate deviationsABS-RATE-DEPT
Overtime Dept Anomalies (ML)Per-unit overtime deviationsWT-OVERTIME (per unit)

Payroll Anomalies (ML) additionally supports Use Calendar Regressor: with it the working-day calendar travels with the detection, so a payroll dip in a short month (say, 19 working days in January) is explained by the calendar instead of being flagged as an anomaly. Same requirements as for forecasts: ARIMA/AutoARIMA model and a generated payroll calendar.

Bradford Factor — responsible use​

The Bradford Factor scores frequent short absences (B = S² × D over a rolling 12 months, where S = distinct sick episodes and D = total sick days) and lists employees above a configurable threshold (default 500). Continuation documents of one electronic sick-leave case (the same certificate case number) count as a single episode, so a single illness posted as several documents does not inflate the score.

This is an occupational-health conversation tool to find people who may need support — not a sanction or disciplinary instrument. Frequent short absences should prompt a wellbeing conversation, not a penalty. The analytic ships disabled with person-level results off; viewing its rows requires the person-level permission.

Privacy​

Forecast and anomaly services receive only anonymized {period, value} pairs — never names or personnel records. Person-level anomaly detail (individual pay jumps, the Bradford list) is shown only when person-level results are explicitly enabled on the instance and the viewer has the person-level permission.

See Also​

© 2008 - 2026 SMART business