Short answer
Predictive HR analytics should estimate a defined workforce quantity for a real planning decision, not claim to know an employee's intentions. Set the unit, outcome and time horizon; compare a simple baseline with any model; backtest on later periods; show uncertainty and assumptions; and plan actions at an appropriate group level under accountable human review.
Start with the decision the forecast must inform
“Predict turnover” is not a decision. “Set recruiting capacity for service roles over the next quarter” is. The second statement identifies an owner, a population, a period and a practical response.
Complete a forecast brief before selecting a method:
| Field | Buyer entry |
|---|---|
| Decision | The staffing, skills, scheduling or budget choice to be made |
| Decision owner | Person accountable for that choice |
| Unit | Role family, site, region or another justified group |
| Outcome | Exact quantity to estimate |
| Horizon | Week, month, quarter or other operating period |
| Available evidence | Data that exists before the forecast date |
| Baseline | Simple comparison the model must improve on |
| Error cost | Operational effect of estimating too high or too low |
| Review date | When actual results and assumptions will be checked |
| Excluded use | Individual label, employment decision or causal claim outside scope |
The 2025 UK government AQuA Book treats analytical quality as a lifecycle responsibility. It distinguishes verification, whether analysis meets its specification, from validation, whether it meets user needs in the intended environment. It is government guidance, but the distinction is useful for any workforce forecast.
Define the quantity before choosing variables
Similar labels can hide different outcomes. “Hiring need” might mean approved vacancies, recruiter workload, accepted offers, people who start, or productive capacity after training. “Turnover” can include or exclude transfers, redundancies, retirements, seasonal contracts and people who leave during the period.
Record:
- the numerator and denominator;
- inclusion and exclusion rules;
- the date each record becomes known;
- how transfers and changing headcount are handled;
- whether the estimate is a count, rate or probability;
- how a later actual result will be calculated on the same basis.
Do not combine different questions because they share data. Forecasting aggregate starts for workforce planning is different from classifying a person as likely to depart. For this aggregate planning job, define outputs without an individual risk list.
Build a baseline before a complex model
A useful model must improve a relevant decision, not merely produce a more elaborate number. Start with a transparent baseline such as:
- the recent average under comparable periods;
- the same period last year where seasonality is relevant;
- approved workload divided by documented productive capacity;
- committed starts, known vacancies and planned role changes;
- a range based on explicitly different operating assumptions.
Compare the candidate method with that baseline on periods neither method could see in advance. A model that fits historical records but does not improve a later planning decision adds cost without useful evidence.
The AQuA Book says uncertainty is inherent in analytical inputs and outputs and should be addressed through the analytical lifecycle. The cross-government Uncertainty Toolkit frames the work as agreeing how uncertainty informs the question, identifying it, assessing it, and communicating it.
Separate forecast, scenario and target
These three numbers serve different purposes:
| Label | Meaning | Suitable use |
|---|---|---|
| Forecast | Method-based estimate under stated assumptions | Plan likely capacity and resources |
| Scenario | Plausible condition, without a claim that it will occur | Prepare contingent actions |
| Target | Desired result chosen by leaders | Set accountability and resources |
If the target is fewer than ten voluntary departures, that does not make ten the forecast. If a severe scenario assumes twenty departures, that does not mean the model assigns that outcome a measured probability.
Show the central estimate with assumptions and a useful range. State whether the range is a formal prediction interval, a sensitivity range, or a set of scenarios. Do not call three management guesses a confidence interval.
Backtest the whole decision process
Choose historical cut-off dates and recreate what would have been known then. Produce the forecast, compare it with what later occurred, and calculate the resulting error and operating cost under declared assumptions. A backtest cannot prove what an unobserved alternative intervention would have caused.
Review more than average error:
- error by time period and relevant workforce group;
- systematic overestimation or underestimation;
- performance during reorganisations, seasonal peaks and policy changes;
- missing or revised records;
- the effect of workforce composition changes;
- whether planners understood and used the uncertainty correctly;
- whether the action created avoidable cost or left capacity uncovered.
The NIST AI Risk Management Framework Core recommends test and evaluation methods tied to context, documented uncertainty, error reports and input from affected and domain experts. Its framework is voluntary and cross-sector. It does not certify an HR model.
Use a forecast register
Keep one reviewable record for every published forecast:
| Field | What to retain |
|---|---|
| Version | Data cut, code or formula, assumptions and owner |
| Baseline comparison | Result from the agreed simple method |
| Forecast | Estimate, range, unit and horizon |
| Scenario difference | Assumption changed and operational implication |
| Known limits | Missing groups, data changes and conditions not represented |
| Decision | Capacity or resource choice made by a person |
| Actual | Later result calculated under the original definition |
| Error review | Direction, size and practical consequence |
| Change | Continue, revise, narrow or stop |
Do not silently update the historical forecast after actual results arrive. Preserve what the decision owner saw at the time.
A fictional aggregate forecast
Northbridge Services is a fictional US and UK employer planning one quarter for a service-role population of approximately 400 people. During the prior 12 months, it recorded 48 voluntary departures under a stable definition, an average of four per month and roughly 1% of average monthly headcount.
An unchanged-rate baseline estimates 12 departures over the next three months: 400 multiplied by 1%, multiplied by three. This is an arithmetic baseline, not a claim that each employee has the same probability or that past conditions will continue.
The team records three planning scenarios:
| Scenario | Assumption | Departures used for capacity planning |
|---|---|---|
| Lower | Recent improvement continues | 8 |
| Baseline | Recent monthly rate remains similar | 12 |
| Higher | A known contract transition creates more movement | 18 |
These are fictional planning assumptions, not a measured probability range. Leaders also want six additional filled roles by quarter end. With no opening vacancy, the baseline capacity plan therefore needs 18 starts: 12 replacements plus six for growth. The lower and higher scenarios imply 14 and 24 starts respectively.
Recruiting, finance and operations agree which preparation is reversible. They reserve assessment capacity for the baseline, identify an approved contingency for the higher scenario and review actual departures, starts, accepted offers and start-date delays under separate definitions. They do not increase recruiter targets by hiding unpaid work or treat internal moves as new group headcount.
At quarter end, they compare the original forecast with actual results and examine which assumptions changed. A good numerical result could still be accidental. A large miss may reflect a method problem, a changed operating condition, or both.
Add qualitative evidence without turning it into a predictor
Employee accounts can help planners identify assumptions to test. For example, repeated descriptions of certification delays may prompt a separate capacity scenario and an operational record check. The accounts do not establish how widespread the issue is, whether it caused departures, or which person will leave.
Use qualitative people analytics when the decision needs mechanisms and context. Use the employee retention strategy guide when the evidence points to an owned change in pay, work, management, development or onboarding.
Where Lontra fits
When planners need current examples behind a defined group-level question, explore Lontra's employee conversations and manager briefs. Managers receive a brief rather than raw employee conversations, and people decide the response.
Lontra does not produce a workforce forecast, assign employee risk, predict departures or make automated employment decisions. Its conversation evidence should remain separate from any claim about prevalence, cause or an individual's future action.
Frequently asked questions
What is predictive HR analytics?
Predictive HR analytics uses historical workforce and operating data to estimate a defined future quantity, such as aggregate hiring demand or workforce capacity, over a stated period. The forecast should show assumptions, uncertainty, validation evidence and the human decision it informs.
Should predictive HR analytics identify who will leave?
An aggregate planning forecast does not need to label individuals. Plan capacity by a suitable group and time period, then address work concerns through transparent conversations and normal people processes rather than treating a probability as an employee fact.
What is the difference between a forecast and a scenario?
A forecast estimates an outcome under a method and stated assumptions. A scenario asks what the organisation would do under a plausible condition. A target states what leaders want. Keep all three labels separate so a desired number is not presented as evidence.
