Predicting Who Will Produce vs Who Will Stay
Hiring teams often treat "will they produce" and "will they stay" as the same question. They are not, and a model that is good at one is not automatically good at the other. A study of 10,765 hires connected screening to production and showed what predicts producing. Whether the same signals predict staying is a separate question, and one the same method can answer once termination data is connected. For roles with high early attrition, you want both lenses, not one standing in for the other.
Source: "Decision Traces," Saad Bin Shafiq, NODES, 2026. Read it on arXiv.
Two different questions
Production asks whether a hire will reach real output, and how much. Retention asks whether they will still be there in a year or three. A candidate can produce well and leave early, or stay for years and never produce much. Optimizing for one does not guarantee the other, so a hiring process needs to be clear about which it is measuring.
What the research shows about production
This part is measured. In the study, resume keywords did not predict production, prior experience and a license were anti-predictive, and personality assessment was the strongest single signal. Speed to production followed a measurable economic constant. These are findings about who produces, grounded in connected ATS and HRIS data. See the research.
What is still open about retention
Retention is the next question, and it is honest to say it is not yet answered here. The study did not include termination data, so it could not measure who stays. The same decision-trace method extends to retention once that data is connected, and the paper sets this up as the next analysis: testing whether the behavioral score predicts retention, and whether agents who ramp faster also stay longer. If they do, the economic case for the infrastructure roughly doubles. That is a hypothesis the method is built to test, not a result NODES has published.
Why both lenses matter
In insurance, most agents leave early. LIMRA puts four-year retention near 15%, with the heaviest losses in years one and two. A producer who leaves in year two is a different and expensive problem than a candidate who never produced. Seeing only one of the two questions leaves money and risk unmeasured. See insurance hiring.
How a decision trace covers both
| Question | What it asks | What connected data shows | Status |
|---|---|---|---|
| Production | will they reach real output | which signals predicted production | measured in the study |
| Retention | will they stay | whether those signals also predict staying | the next analysis, once termination data connects |
A decision trace connects screening inputs to outcomes, so the same structure that answered the production question can answer the retention question as the data arrives.
Frequently asked questions
Is predicting production the same as predicting retention? No. They are different outcomes, and a signal that predicts one may not predict the other. They need to be measured separately.
Does NODES predict employee retention? The published study measured production, not retention. The method extends to retention once termination data is connected, and the paper frames that as the next analysis rather than a current result.
Why does retention matter so much in insurance? Most agents leave early. LIMRA puts four-year retention near 15%, with most departures in years one and two, so early attrition is a major cost.
Can one system answer both questions? Yes, in principle. A decision trace connects screening to outcomes, so the same structure can test production now and retention as termination data becomes available.
Related reading
- Prediction vs moderation
- The speed-to-production constant
- AI hiring intelligence for insurance carriers
See what your own data predicts about production today, and what it could show about retention. Book a 30-minute walkthrough.