Workforce Analytics Architectures for Predicting Attrition and Capacity Gaps in Large Enterprises

Abstract

Attrition prediction is usually presented as a modelling problem. In a large enterprise it is first an architecture problem. The signals
that carry information about whether a person will leave are distributed across a human capital management system of record, a
performance module, a talent and career interest module, a learning platform, system access logs and purchased labour market data.
None of those systems was designed to be joined to the others, they refresh on incompatible cycles, they carry different identifiers
for the same person, and they record events at a different time from when the events occurred. This paper is a practitioner
architecture account of a workforce analytics platform built across six such sources for a workforce of about fifteen thousand people
in more than forty locations. It describes the identity resolution layer that makes any downstream analysis possible, the provenance
and refresh cadence of each feature, and what breaks when a source changes. It then argues that the analytically interesting target is
not attrition but regretted attrition, because predicting departures is easy and predicting the ones that hurt is the actual problem:
reported figures show total attrition close to flat near twelve percent while the regretted share of departures fell from twenty two
percent to eight percent over four years. The paper treats a flight risk score as a governed artefact rather than a measurement, sets
out what such a score is and is not licensed to do, and reports the departure model's evaluation exactly as described, including one
case where the reported precision does not reconcile with the reported confusion matrix. Capacity gap forecasting is presented as
the operational output that makes prediction actionable, and internal mobility as the intervention lever rather than a reporting
dimension. Six signals the organisation found predictive and six it found weak are reported with their conditional rates. All figures
are the author's own operational reporting from a single enterprise, and no controlled evaluation was performed.