Compensation flight-risk model

A compensation flight-risk model surfaces people whose performance rating, current pay, and approved market vintage sit in a combination compensation already treats as a retention risk. It does not change pay. It does not forecast who will resign. It does not invent an attrition percent or a market percentile. The output is a flag with cites, plus a recommended adjustment that still needs someone in compensation to approve.

HRBPs use the flag to open a conversation with the manager and with compensation, not to promise a number. Compensation uses it to decide whether an off-cycle or in-cycle adjustment is warranted against the same rules that already govern pay. If a cite is missing, that field stays empty. An empty cite is not a reason to fill the gap with a rounded number.

This is narrower than predictive attrition forecasting. Attrition models score likelihood of leaving from many signals. This model only asks whether a high performer is paid below the approved benchmark vintage, with enough history to explain how they got there. Mixing the two produces a score that looks urgent and cannot be audited.

Do not run the model until three inputs are loaded and dated.

Performance rating comes from the most recent completed cycle, with the cycle name and close date attached. Informal manager comments, mid-year notes, and on-track labels are not ratings. If the person is new to the rating scale, or the last cycle was skipped, the rating field stays empty.

Current pay is base, plus any other cash element your policy includes in the comparison, as of a named payroll effective date. Do not mix a pending offer or unposted merit into base unless that is how your comparison is defined. Pull pay from the HRIS of record. Workday is a common source. Other HCM systems work the same way: one extract, one effective date, no spreadsheet overlay.

The approved benchmark vintage is the survey cut compensation has already signed off, not the latest file a vendor sent last week. Radford and CompAnalyst are typical sources of that vintage. The model must store the survey name, the job match, the geography, and the effective date of the cut. Live market benchmarking can refresh that cut on a schedule. This model does not invent a new cut on the fly.

Pay history vintage is the dated trail of prior bases and prior increases that explain the current position. It is a list of effective dates and amounts already in the system, not a narrative.

People-analytics platforms such as Visier often hold the joined view of rating, pay, and tenure. Treat those platforms as an assembly layer. They do not approve the benchmark vintage. Compensation does.

If any of the three required cites cannot be loaded, do not invent a substitute. Leave the field blank and do not raise a complete flag.

A complete flag cites three things: the performance rating with its cycle, current pay versus the approved benchmark vintage (survey name, match, geography, and vintage date), and the history vintage (prior effective dates that explain how pay arrived here).

The comparison is below the approved vintage for this match, not below a percentile someone prefers. Do not compute or display a market percentile unless that percentile is already a defined field on the approved vintage and compensation has authorized it as the comparison rule. If the vintage does not carry that field, leave it empty. Do not back-fill a midpoint because the model usually uses one.

The same rule applies to attrition. Do not attach a resignation probability, an unexplained flight-risk score, or a team attrition rate. If someone wants a likelihood-to-leave view, that is a different model with different inputs.

Here is an illustrative path, not a measured outcome. A senior individual contributor in a matched engineering job has an Exceeds rating from the latest closed cycle. Current base, as of a named payroll date, sits below the approved Radford vintage for that job and location. History shows the last increase on a named date and no off-cycle since. The flag cites those three facts and recommends that compensation consider an adjustment under existing policy. It does not name a target percentile or say the person will leave. If that vintage is missing from the extract, the pay-versus-market field stays empty and there is no complete flag.

When a field is empty, show empty. Do not display a placeholder that reads like a value, and do not grey in an estimate. Reviewers must see what is missing.

A recommended adjustment, if any, is a pointer to the policy path that already exists: off-cycle retention, in-cycle merit, or no action until the next cycle. The suggestion is not posted pay.

Compensation owns the decision. The model does not write to payroll, does not create a compensation change that auto-completes, and does not email the employee.

The review packet should include the three cites, empty fields, recommended policy path, and requester (HRBP or manager). Comp checks that the job match still holds, that the vintage is the approved one, that the rating is the official cycle rating, and that history has not omitted a recent increase. Comp also checks continuous pay-equity monitoring so a retention move does not create a new inequity.

If comp approves, the change follows the same workflow as any other pay change: approval chain, effective date, and employee communication owned by the manager and HRBP. If comp declines, the flag stays visible as a declined recommendation with a reason, so the same person is not re-flagged next week on identical inputs.

Do not treat a flag as a raise. A manager who tells the employee the model will adjust them has already spent budget that does not exist. The flag is internal. The employee conversation happens only after approval, and the number is the approved number.

The merit cycle agent is the in-cycle counterpart. Flags that land near cycle lock should fold into cycle recommendations rather than land as a second increase on top of merit. Off-cycle use is for cases policy already allows between cycles.

Three failure modes show up quickly if you do not name them.

A flag with no vintage. Someone loads last year's survey file, a recruiter range, or a quick CompAnalyst pull that compensation has not approved. The model then compares pay to a number that is not the company benchmark. The flag looks precise and is wrong. Hard rule: no approved vintage date, no complete flag. Empty stays empty.

Treating the flag as a raise. Operations teams want a closed loop: flag, then auto-submit an increase, then notify the manager. That loop bypasses match review, equity review, and budget. It also invents an amount the model was never allowed to set. The model may recommend that compensation consider an adjustment. It may not pick the amount, post it, or tell the employee.

Inventing an attrition percent. Do not park a resignation probability or a team attrition rate next to the pay flag. A separate attrition model, if you run one, keeps its own methodology and must not be blended into this flag. This page cites rating, pay versus approved vintage, and history vintage. That is the whole claim.

A quieter failure is using this flag as a substitute for equity analysis. Below-vintage high performers are one slice. Equity monitoring will surface people who are not high-rated and still underpaid relative to peers. Do not wait for a flight-risk flag to fix those.

Market work defines the vintage this flag is allowed to use. Equity work checks the side effects of any approved adjustment. Attrition forecasting, if you run it, is a separate score with separate governance. Cycle automation applies approved rules at scale when the annual process is open.

Vendors in this stack are sources and assembly layers: Workday for HRIS pay and job data, Visier-class people analytics for joined views, Radford and CompAnalyst for survey vintages. None of them is the approver. Compensation remains the approver. The model remains a cited flag. Pay changes only when a person says so.