Big data analytics in HR promises prediction. The constraint most employers hit first is unglamorous: their own data is small, inconsistent across systems, and mostly about people who are still there.
Most employers do not have big data about their people
A workforce of a few thousand with a handful of years of history is a small dataset by any statistical standard, and the events you want to predict, such as resignations in a specific role, are rarer still. That does not make analysis useless; it means confidence intervals are wide and conclusions drawn from a few dozen cases should be held loosely.
Consistency across systems is the real blocker
Recruitment, HR, payroll, learning and engagement data usually live in different systems with different identifiers and different definitions of a job. Joining them reliably is the majority of the work in any analytics project. Employers who skip it get fast answers built on mismatched records, which is worse than no answer because it is confident.
Correlation in workforce data is especially treacherous
Almost every interesting workforce variable is confounded by tenure, role, location and manager. A finding that people who attend training leave less often is usually a statement about who gets offered training. Before acting, ask what else changed with the thing you measured, and prefer a small controlled comparison over a large uncontrolled one.
There is a hard line around decisions about people
Using a model to prioritise attention is one thing; using it to make or substantially drive decisions about hiring, pay or exit is another, and it attracts scrutiny under both employment and data protection rules. Keep a person accountable, be able to explain the basis of a decision in plain terms, and test outputs for disparate impact before they are used.
Questions people ask about big data analytics hr
Is predictive attrition modelling worth doing?
Sometimes, at scale, and only if somebody will act on it. At smaller sizes the honest answer is usually that managers already know.
What is the first thing to fix?
Identifiers and job definitions across systems. Everything else depends on being able to join records correctly.
Can we use a model to screen candidates?
Only with great care. Selection procedures are regulated and have to be validated and monitored for adverse impact.