An Employee Turnover Prediction Model with “85% Accuracy”: Why That Number Can Mislead, and What to Look At

Imagine a vendor presenting to the HR team. “Our turnover prediction model has 85% accuracy.” The room nods. Sounds impressive, doesn’t it? Eighty-five out of a hundred predictions are right. The contract is nearly signed.

Before that happens, ask one simple question: “What percentage of employees at our company actually resign each year?”

If the answer is 15%, there is some unwelcome news. A model that always guesses “nobody will resign” will be right for 85% of employees without learning anything. Its accuracy is 85%. And that model is completely useless, because it never finds a single employee who is going to resign.

This is one of the most common traps in machine learning projects for HR, and in other fields where the event is rare: accuracy can be high precisely when the model does not work. Let us take it apart so you can assess such claims critically.

The confusion matrix: four possibilities

Every binary prediction (will resign or will stay) has four possible outcomes:

• True positive (TP): the model predicts resign, and the person does resign.
• False positive (FP): the model predicts resign, but the person stays.
• True negative (TN): the model predicts stay, and the person does stay.
• False negative (FN): the model predicts stay, but the person resigns.

From those four numbers we compute several measures:

• Accuracy = (TP + TN) / all predictions. How often the model is right overall.
• Precision = TP / (TP + FP). Of everyone predicted to resign, how many actually resigned.
• Recall (sensitivity) = TP / (TP + FN). Of everyone who actually resigned, how many the model managed to find.

An example with numbers

Suppose a company has 1,000 employees, and 150 of them resign this year (15%).

Model A: “nobody resigns”. Its predictions are right for the 850 who stay, and wrong for the 150 who resign.

• Accuracy: 850/1,000 = 85%
• Recall: 0/150 = 0%
• Precision: undefined (it never predicts resignation)

Model B: a model that really learns. It flags 140 employees as high risk. Of those 140, 90 do resign (TP = 90), and 50 stay (FP = 50). Of the 150 who resign, it misses 60 people (FN = 60). So TN = 850 − 50 = 800.

• Accuracy: (90 + 800)/1,000 = 89%
• Precision: 90/140 = 64%
• Recall: 90/150 = 60%

Notice: the accuracy of the two models is not far apart (85% and 89%), even though one is entirely useless and the other finds 60% of the employees who will resign. Accuracy hides the most important difference. That is why, when someone offers a single accuracy figure, ask for other measures.

Which measures should you ask for?

It depends on the cost of errors:

• If losing a key employee is expensive and you want to find as many at-risk people as possible, recall matters. The consequence is that precision may drop (more false alarms).
• If intervention is expensive (for example an individual retention program with a high cost), precision matters so resources are not wasted on people who were not planning to leave.
• For a picture of the balance: the F1-score (the harmonic mean of precision and recall) and AUC-ROC (the model’s ability to distinguish those who resign from those who stay across different thresholds) are often reported.

And one important question that is often forgotten: compared with what? Is the model better than a simple rule, for example “employees with under two years of tenure and no promotion in 18 months are high risk”? A complex model must prove its added value over a simple baseline.

Five other traps in HR prediction projects

1. Data leakage. A model that accidentally uses information that only becomes available after the event, for example “has submitted a resignation letter” or “final contract status”. Results on test data look extraordinary, but in the real world the model is useless because that information does not yet exist when the prediction is needed.

2. Testing on the same data used for training. A model must be tested on data it has never seen, ideally data from a period after the training data, so it reflects the real situation.

3. Unrepresentative data. A model trained on data from one division does not necessarily apply to another. Changing conditions (for example a shifting job market) also make models go stale quickly, so they need to be monitored and retrained.

4. Ignoring explanation. A “high risk” prediction without reasons is hard to act on. HR needs to know what factors drive the risk (for example time since last promotion, workload, or commute distance), not just a score. Model interpretation techniques help here.

5. Ethics and privacy. Predictions about individuals involve employees’ personal data. Ask: do employees know their data is being used? Does the model treat certain groups unfairly (for example based on age, gender, or family status)? Will the risk score be used to help, or to punish? A prediction model should be a tool to support employees and improve the work environment, not to surveil or push people out.

From prediction to action

A prediction is useless if no action follows. A healthy flow usually goes like this:

1. Define the business problem: which group’s turnover costs the most?
2. Prepare clean and relevant data, with attention to privacy.
3. Build a simple model first as a baseline, then a more complex one if needed.
4. Evaluate with measures that match the cost of errors, not just accuracy.
5. Translate results into action: conversations with supervisors, workload adjustments, career paths, or development programs.
6. Monitor and update: measure whether the intervention works, and update the model regularly.

In the end, the goal of the model is not to produce a pretty score. Its goal is to help the company retain the people worth retaining in a fair and humane way.

Seven questions for your vendor or internal team

If you are the party buying or commissioning a prediction project, ask these seven questions: (1) What is the proportion of the predicted event in the data, and against what baseline is the model compared? (2) Which measures are used besides accuracy? (3) How was the data tested, and is there data from a different period for testing? (4) Which features are used, and could any leak or be unethical? (5) How are predictions explained to users, so HR knows the reasons? (6) How is the model monitored and updated? (7) What concrete action will be taken based on the predictions, and how will we measure whether the action works? If any question cannot be answered, there is homework to do.

A pilot before full rollout

An accurate prediction does not necessarily produce better retention. The only way to be sure is a controlled pilot: choose several groups of high-risk employees, apply an intervention to some of them (for example a career development conversation), and compare with those who do not receive the intervention, with random assignment if possible. If the resignation rate in the intervention group is meaningfully lower, we have evidence that the combination of prediction and action works. Without a test like that, we only have a pretty risk score and good intentions.

Short Q&A

Does a resignation prediction model violate employee privacy? It can, if done without a clear policy. Healthy practice includes a legal basis and consent in line with personal data protection rules, restricted access to individual scores, using as little data as possible, and transparency to employees about the purpose of use.

How much data is needed to build a model like this? There is no single number. What matters is having enough resignation events for the model to learn from, not just the number of employees. For small companies, descriptive analysis and exit interviews are often more useful than a fragile predictive model.

---