Understanding Your Time to Separation Results

Your Time to Separation model has finished running. Now the important part is understanding what the results actually mean. The Results Summary brings together the key information you need to evaluate the model: how well it performed, which factors are associated with higher or lower separation risk, how risk changes across your selected time horizons, and what the model predicted for the population you scored.

This guide walks through the report section by section and explains what to pay attention to, what the main measures mean, and which results should give you confidence or prompt a closer look.

Unlike other One AI model types, Time to Separation does not have a separate Exploratory Data Analysis (EDA) report. Everything you need to review the model is included in the Results Summary.

Finding the Results Summary

1. Select One AI from the top navigation bar.
2. Find the Time to Separation model you want to review and select Runs.
3. Select the run you want to review.

A completed Time to Separation run opens directly to the Results Summary rather than an EDA report.

Key terms used in the report

• Separation risk: The predicted probability that an employee will have separated by a particular time horizon.
• Survival probability: The predicted probability that an employee will still be active at that same time horizon.
• Prediction horizon: A future point in time from the Prediction Start Date when separation risk is estimated, such as 12, 24, or 36 months.
• Risk percentile: The employee’s relative risk rank within the scored population. The value is stored from 0 to 1. For example, 0.99 means the employee has higher predicted risk than about 99% of the scored population.
• High-risk flag: Indicates that an employee meets the model's selected higher-risk threshold for that timeframe.
• Restricted Mean Survival Time (RMST): The employee’s expected active days within a selected window. It is not a predicted separation date.

Risk and survival are two views of the same prediction at the same horizon: Risk + Survival = 1. For example, 0.25 separation risk corresponds to 0.75 survival probability.

Executive Summary

The Executive Summary is the best place to begin. It gives you the quickest read on how the model performed and highlights the most important results from the run.

Key Takeaways

Key Takeaways gives you the shortest summary of the run. Start here before moving into the more detailed tables and charts.

Model Performance at a Glance

Model Performance at a Glance summarizes how well the model ranked employees by risk, how well it distinguished between employees who experienced a separation within a timeframe and those who did not, and whether there was enough historical data to evaluate the longest timeframe.

Start with the Takeaway column. The ratings are designed to help you quickly spot areas that look strong and areas that may need a closer look.

No single rating determines whether the model is useful. Review the results together, especially when the longest timeframe has limited historical support.

Core Findings

The Core Findings section gives you more detail about how the model performed, how separation risk differs across employee groups, which signals mattered most, and what the predictions look like for the current scored population.

Run at a Glance

Run at a Glance confirms the main settings behind the predictions in the report:

• Prediction Start Date: The date the model starts predicting from.
• Risk Timeframes: The future points where separation risk is estimated.
• Expected Active Days Window (RMST): The window used to calculate expected active days.

Check these values first to make sure you are reviewing the run you intended.

How This Run Performed

This section gives you more detail about the model’s historical performance and whether there was enough history to evaluate each timeframe.

The Main Validation Timeframe (Evaluation Horizon) is the primary timeframe used to judge model performance during validation.

Performance by Timeframe compares performance across each timeframe. Pay particular attention to Validation periods with enough history. A longer timeframe may not have enough historical follow-up to evaluate even when the shorter timeframes do. If Event Accuracy or Risk Accuracy shows "Not available", that usually means the timeframe could not be fully evaluated from the available history rather than that the model received a poor score.

Validation Period Overview shows how ranking performance changed across the historical periods used to test the model. Look for whether the results are reasonably consistent from one period to another.

How Well Predicted Risk Matched Actual Outcomes

This chart compares the model’s predicted risk with what actually happened in historical data.

Use the Takeaway directly below the chart as the report’s summary of the overall pattern for this run.

How Risk Differs Across Employee Groups

This section compares lower-risk and higher-risk groups and shows how those differences develop over time.

The Risk Group Summary divides the scored population into five groups, from the lowest predicted risk to the highest. It lets you compare predicted risk, group size, and Expected Active Days across those groups.

Larger differences between the groups mean the model is finding clearer differences across the population.

Expected Active Days (RMST) is not a predicted separation date. It summarizes expected active time within the selected RMST window.

The Event Risk Over Time by Risk Group chart shows separation risk building over time. Higher lines represent groups whose predicted separation risk increases more quickly.

The Probability Still Active Over Time by Risk Group chart shows the same predictions from the opposite perspective: the probability that employees remain active as time passes.

Look at the spacing between the groups in both charts. Larger gaps mean the model sees more meaningful differences between lower-risk and higher-risk employees.

Main Signals in This Run

This section helps you understand which employee attributes had the strongest relationship with separation risk in this model.

Feature Importance

The model considers many attributes when it estimates risk. Use the Feature Importance table to see whether each feature is associated with higher or lower risk and how strong that relationship is.

Start with Signal Direction and Effect Size (Hazard Ratio). Values farther from 1.00 represent a stronger relationship with risk. The Model Weight (Coefficient) is there for more technical review.

These are associations found by the model, not proof that a feature causes an employee to separate.

Use the Takeaway below the table for the report’s overall summary of the signals in this run.

Current Population Snapshot

The Current Population Snapshot shifts from historical model validation to the employees who were actually scored in this run.

Current Population Summary

Current Population Summary confirms how many employees were scored and the Prediction Start Date used for those predictions. Check these values before interpreting the rest of the population results.

Predicted Risk by Timeframe

Predicted Risk by Timeframe shows the average and median predicted separation risk for the current population at each selected timeframe.

Risk will often increase at longer timeframes because there is more time in which a separation could occur. Use the Takeaway below the table to understand the overall pattern in this run.

Employees Flagged as Higher Risk

Employees Flagged as Higher Risk shows the selected threshold for each timeframe and how many employees met it.

The threshold is determined separately for each timeframe. If the report does not have enough historical evidence to support a threshold, the threshold and flag results will show as unavailable.

The notes below the table explain why a threshold was or was not published for each timeframe.

Baseline Time-to-Risk Policy

Baseline Time-to-Risk Policy provides additional detail behind the higher-risk thresholds.

Most users only need to check Application Status and, if the policy is unavailable, the Unavailable reason.

If the policy cannot be established from the available historical data, higher-risk flags cannot be published for that run.

Technical Appendix

The Technical Appendix contains additional detail about the model setup, validation, and how the data was prepared. Most users will not need to review every table here. 

Run Setup

Run Setup contains the full technical settings used for the run. It repeats some information from Run at a Glance and adds settings used for validation, model selection, and the time-to-risk policy.

For a standard review, the most important fields are the Prediction Start Date, Risk Timeframes, Expected Active Days Window, and Main Validation Timeframe. The remaining fields are useful when comparing runs or troubleshooting with One Model Support.

Detailed Validation Results

Detailed Validation Results shows performance separately for each validation period and timeframe.

Use this section when you need to understand why a result was unavailable, investigate a timeframe that could not be fully evaluated, or compare performance across historical periods.

Feature Model Full List

Feature Model Full List shows every feature used in the final model. Use Feature Importance in Core Findings for the quicker view of the strongest signals. Use this full list when you want to see everything the final model relied on.

Feature Preprocessing and Selection

Feature Preprocessing and Selection shows how the attributes supplied to the recipe were prepared before the final model was built.

Summary Counts gives you a quick view of how many features were supplied, generated, removed, and ultimately used.

Input Features Supplied shows the attributes that entered the modeling process and their final status.

Preprocessing Applied documents the preparation steps One AI performed before modeling.

Features Removed Before Modeling shows which features were removed and why. Check this table if a feature you expected to see in the final model is missing.

A removed feature does not automatically indicate a problem. This section is primarily there to make the model-building process transparent.

Person-Snapshot and Outcome Assumptions

This section documents assumptions used when building the model. Each eligible employee at a point in time is treated as one observation, which allows the same employee to be scored at different points in time. The model also assumes that separation records are complete through the available outcome date. Most users will not need to take action here, but this information can be useful when validating the underlying data.

Results File Preview

This section shows a sample of the employee-level results available in the downloadable file. For each configured timeframe, the results can include:

• Survival Probability: The predicted probability the employee is still active by that day.
• Event Risk: The predicted probability the employee has separated by that day.
• High-Risk Flag: Whether the employee met the higher-risk threshold for that timeframe.
• Risk Percentile: Where the employee’s predicted risk ranks relative to the rest of the scored population, stored from 0 to 1.

Risk Percentile and Event Risk answer different questions. Event Risk is an estimated probability. Risk Percentile tells you how the employee ranks compared with the other employees being scored.

The report only shows a preview. Download the results CSV to view the full employee-level output.

Putting the Results Together

You do not need every result in the report to be perfect. The goal is to understand the model well enough to know where it is strong, where it has limitations, and whether those limitations matter for how you plan to use the predictions.

Before moving forward, make sure the model was evaluated on the timeframes you care about, the risk patterns and main signals make sense for your workforce, and you understand any warnings or unavailable results.

If something does not look right, revisit the model configuration or ask for help before deploying. If the results are clear and appropriate for your use case, you can move on to deploying the model and using the predictions in One Model.

Ask your Customer Success team about the accompanying Time to Separation Storyboards. They can help you turn the model outputs into a clearer story for stakeholders, explore the results in more detail, and identify where the insights may support action.

For instructions on deploying the model, see Creating a One AI Time to Separation Model.