Results
Calculated results appear here after analysis.
Cumulative Gain Chart
Selected Cutoff Analysis
Interpretation
Model Comparison
| Model | Gain @ 10% | Gain @ 20% | Gain @ 30% | Gain @ 50% | AUGC | Normalized Gain | Max Separation |
|---|---|---|---|---|---|---|---|
| No model results yet. | |||||||
Segment Comparison
| Segment | Rows | Positive Rate | Gain @ 20% | Lift @ 20% | Normalized Gain |
|---|---|---|---|---|---|
| No segment results yet. | |||||
Gain Summary Table
No rows calculated.| Bucket | Score Range | Population | Population % | Cumulative Population | Cumulative Population % | Positives | Cumulative Positives | Positive Rate | Cumulative Gain | Random Gain | Gain Above Random | Bucket Lift | Cumulative Lift | Minimum Score |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| No gain table yet. | ||||||||||||||
Formula Used
Cumulative Gain (%) = Cumulative positive weight captured ÷ Total positive weight × 100.
Population (%) = Cumulative selected weight ÷ Total observation weight × 100.
Cumulative Lift = Cumulative gain percentage ÷ Cumulative population percentage.
Normalized Gain compares the model area against random and perfect ranking areas.
How to Use
- Paste or upload data with actual labels and prediction scores.
- Enter the actual label column and one or more model score columns.
- Choose the positive class, grouping method, cutoff, and chart settings.
- Select Analyze Data to build curves and detailed tables.
- Review gains, lift, thresholds, segment performance, and export options.
Example Data Format
| id | actual | model_a | model_b | weight | segment |
|---|---|---|---|---|---|
| 1 | 1 | 0.96 | 0.82 | 1 | Retail |
| 2 | 0 | 0.89 | 0.76 | 1 | Retail |
| 3 | 1 | 0.85 | 0.91 | 1 | Enterprise |
| 4 | 0 | 0.72 | 0.61 | 1 | Enterprise |
Interpretation Guide
A curve above the diagonal baseline indicates useful ranking performance. A steeper early rise means the model captures positives quickly. This matters when only a limited population can be targeted.
The perfect curve shows ideal ordering with every positive ranked first. Gain charts evaluate ranking quality rather than probability calibration. Review the cutoff matching your operational capacity.
Lift compares model capture against random selection at the same population depth. Segment results can expose uneven model performance. Always validate findings on unseen or production-like data.
Frequently Asked Questions
What does cumulative gain measure?
It measures the percentage of all positive cases captured after selecting a percentage of the ranked population.
How is a gain chart different from a lift chart?
A gain chart shows captured positives. A lift chart divides model gain by random expected gain.
Should scores be probabilities?
No. Any numeric score works when its ordering represents model confidence or predicted risk.
Why is the random baseline diagonal?
Random selection captures positives in direct proportion to the selected population percentage.
What is a strong early gain?
A strong early gain captures a large positive share within a small selected population.
Can multiple models be compared?
Yes. Enter several score columns and the calculator plots and ranks every valid model.
How are tied scores handled?
Ties can remain together across bucket boundaries or follow their original stable row order.
What does normalized gain mean?
It scales area improvement from random performance toward the best theoretically possible ranking.
Can weighted observations be used?
Yes. Provide a positive numeric weight column to calculate weighted population and positive capture.