Predictive People Analytics: An Aggregate Forecasting Framework

Predictive people analytics should estimate a defined workforce outcome for a defined group and period, with an uncertainty range and a decision owner. Start with aggregate planning questions, compare against a simple baseline, test on unseen periods, record errors and drift, and keep employment decisions under accountable human review.

Choose an aggregate decision before choosing a model

Begin with a planning decision that does not require a probability for a named employee. Examples include:

• How many voluntary departures should a service function plan for next quarter?
• Which role families may have insufficient skill coverage under the demand scenarios?
• How many hires may be needed by location over the next two quarters?
• Which onboarding process needs investigation because early exits exceed its expected range?
• Which workforce assumptions create the greatest capacity risk?

Write the forecast specification:

The U.S. Office of Personnel Management's workforce planning guidance lists sources such as personnel data, employee input, projected employment trends, separations, hiring, retention, competency assessments, and labor-market trends. The useful unit depends on the decision. More fields do not automatically make a better forecast.

Build a simple baseline first

A complex model should earn its place by improving a relevant decision over a transparent baseline. Depending on the target, compare it with:

1. the last observed period;
2. the same period last year;
3. a rolling average;
4. a seasonal average;
5. the approved workforce plan;
6. a simple rate applied to the eligible population.

Record the baseline before model selection. If a complex model does not improve error, uncertainty, lead time, or decision usefulness, keep the simpler approach.

Accuracy alone is not sufficient. A forecast delivered after the staffing decision has closed may be statistically better and operationally useless. A slightly wider range delivered earlier may support a better capacity plan.

Prepare data around the forecast date

The most common workforce data mistake is using information that would not have been available when the prediction was supposed to be made. Create a data cutoff for every historical forecast.

Your preparation table should include:

Do not place a later exit reason, manager note, or corrected record into an earlier backtest unless it was genuinely available at that time. That leakage makes historical performance look better than a live forecast can be.

Qualitative employee material can help explain a pattern or identify a process hypothesis. It should not be converted into a hidden personal risk score. Preserve source, purpose, coverage, missingness, and reviewer interpretation.

Backtest by horizon and population

Hold out historical periods that the model did not use for fitting. Recreate what the system would have known at each forecast date, then compare the forecast with the outcome.

Track at least:

• absolute error: how far the predicted count is from the observed count;
• directional error: whether the model systematically forecasts too high or too low;
• error by horizon: one-month and two-quarter forecasts should not be mixed;
• error by population: large headquarters groups can hide failure in smaller field groups;
• interval coverage: how often the observed result falls inside the stated range;
• calibration: whether events described as similarly likely occur at similar rates;
• decision error: how often the forecast would have triggered an unnecessary or missed action.

Use enough historical periods to represent seasonality and known operational changes. If reorganizations, acquisitions, policy changes, or missing fields break comparability, show that limit.

NIST recommends defining scope, test methods, knowledge limits, representativeness, validity, reliability, uncertainty, and human roles in its AI Risk Management Framework Core. The U.S. Government Accountability Office organizes its AI Accountability Framework around governance, data, performance, and ongoing review. These are useful questions even when the forecast is a conventional statistical model rather than AI.

Show ranges and scenarios

A single number invites false confidence. Present:

• a central estimate;
• a reasonable uncertainty range;
• assumptions that most affect the range;
• a business-as-usual scenario;
• at least one plausible higher and lower scenario;
• the decision that changes across those scenarios.

Suppose a fictional service group has recorded between 8 and 14 voluntary departures in comparable 12-month periods. Its staffing plan also depends on uncertain demand and a planned role redesign. A forecast for the next quarter might present 3 to 6 departures, not "four people will leave." The planning team can then model recruitment and coverage at both bounds.

The same group hears repeated employee accounts that schedule notice is difficult. That theme can justify a process review, but it does not prove the forecast or identify who will depart. The forecast, employee accounts, and operating records are different evidence types. Leaders can use them together without pretending they say the same thing.

Design the action thresholds before seeing the output

For each range, state the planned action:

Avoid automatic employment actions. A forecast may change a hiring plan, prompt a process investigation, or identify an assumption that needs evidence. It should not decide promotion, pay, discipline, termination, or access for an individual.

Review drift and forecast failures

Choose a review cadence based on how quickly the workforce and decision change. Log:

• forecast and version;
• data cutoff and source versions;
• uncertainty range;
• observed outcome when available;
• error against the simple baseline;
• changes in population, policy, labor market, or collection;
• corrective action and owner;
• decision to continue, restrict, retrain, replace, or stop.

Define acceptable drift and failure conditions before launch. A reorganization may change job families. A new scheduling policy may break the historical relationship. A changed questionnaire may alter a qualitative theme. A model that once performed adequately can become irrelevant.

GAO's framework calls for planned review of performance, acceptable data and model drift, traceable results, and documented corrective action. Treat a forecast as a maintained decision process, not a model delivered once.

Keep individual decisions out of the aggregate forecast

An aggregate target does not remove every risk. Small groups, rare roles, and combined filters can reveal a person. Set minimum useful populations, restrict access, and test whether users can reconstruct identity from role, location, time, or a distinctive event.

For UK processing, the ICO's current guidance on automated decision-making and profiling explains additional rules around solely automated significant decisions and says organizations still need a lawful basis, transparency, accuracy controls, rights processes, and appropriate safeguards. The page notes that it is under review following legislative changes, so confirm current requirements with the responsible legal and privacy advisers.

In the United States, the EEOC's AI and algorithmic fairness initiative states that emerging tools used in employment decisions must comply with the federal civil-rights laws it enforces. A human clicking "approve" is not a meaningful safeguard if that person cannot question the output, inspect evidence, or choose another action.

Use a forecast review card

Give every decision maker one page containing:

1. target, population, horizon, and data cutoff;
2. central estimate and uncertainty range;
3. simple baseline and recent comparative error;
4. main assumptions and missing data;
5. scenario that changes the decision;
6. known limits by population and period;
7. source and model version;
8. prohibited uses;
9. decision owner and review date;
10. next outcome needed for evaluation.

This keeps the conversation on decisions and uncertainty rather than the authority of a score.

Where Lontra may fit

Lontra can be evaluated when workforce planners need reviewed employee context for an aggregate process question. Review the current platform capabilities, then use one focused campaign, up to 30 invitations over 60 days with no credit card, to test collection, source review, and aggregate reporting alongside the forecast process.

That campaign does not create a validated forecasting model or a named-person prediction. Managers receive a brief rather than raw employee responses, and aggregate HR views use a minimum of five respondents. The threshold is a reporting control, not an anonymity guarantee. Keep forecast construction, validation, and every required integration or export as separate verified work.

Sources

• NIST: AI Risk Management Framework Core
• U.S. Government Accountability Office: AI Accountability Framework
• U.S. Office of Personnel Management: Workforce planning and human capital analytics
• ICO: Rights related to automated decision-making and profiling
• EEOC: Initiative on AI and algorithmic fairness