Threshold selection result
The selected cutoff and its classification trade-offs appear below.
Data and threshold statistics
Threshold comparison
Confusion matrices
Main threshold selection curve
Confusion counts across thresholds
Precision–recall trade-off
Sensitivity–specificity comparison
Cost and net value curves
Predicted positive rate
Metric difference curve
Threshold results table
Select a column heading to sort all calculated thresholds.
Formula used
A threshold converts each probability into a positive or negative prediction. Every cutoff produces a new confusion matrix. Those counts determine all displayed metrics.
How to use
- Paste CSV data or load an example preset.
- Confirm the label, probability, and weight columns.
- Choose the threshold range and classification rule.
- Select an optimization method and optional constraints.
- Enter business costs when mistakes have unequal impact.
- Calculate and inspect curves, tables, and comparisons.
- Export the selected result for reporting or deployment.
Example data table
| Actual | Probability | Weight | Label |
|---|---|---|---|
| 1 | 0.92 | 1 | Record 1 |
| 0 | 0.74 | 1 | Record 2 |
| 1 | 0.63 | 1 | Record 3 |
| 0 | 0.21 | 1 | Record 4 |
Interpretation guide
Lower thresholds usually classify more observations as positive. Recall often rises while specificity and precision may fall. Higher thresholds normally reverse that trade-off.
Metric-only optimization suits balanced analytical objectives. Cost-sensitive optimization suits operational decisions with unequal mistakes. Constraints protect minimum business or safety requirements.
Common mistakes
- Selecting a threshold on training data without validation.
- Ignoring class imbalance and predicted positive volume.
- Assuming every false positive has equal business cost.
- Optimizing accuracy when missed positives are critical.
- Using uncalibrated probabilities as reliable risk estimates.
- Deploying a threshold without monitoring distribution shift.
Frequently asked questions
What is a classification threshold?
It is the probability cutoff used to assign predicted classes. Changing it alters positive and negative decisions. It also changes every confusion-matrix metric.
Why is 0.5 not always optimal?
Class imbalance, model calibration, costs, and goals differ. A neutral cutoff may miss important operating requirements. Validation data reveals a more suitable threshold.
Which metric should select the threshold?
Use the metric matching the real decision objective. F1 balances precision and recall equally. Cost minimization handles unequal business consequences directly.
What does Youden’s J measure?
It combines sensitivity and specificity into one statistic. Larger values indicate better separation from random classification. It is common in diagnostic threshold analysis.
When should I use F-beta?
Use it when precision and recall need unequal importance. Beta above one favors recall more strongly. Beta below one favors precision more strongly.
How do constraints affect selection?
Constraints remove thresholds violating required operating limits. Optimization then occurs only among remaining feasible cutoffs. No result appears when none satisfy them.
Why use sample weights?
Weights represent unequal importance, frequency, or sampling design. They produce weighted confusion counts and metrics. This can better reflect deployment populations.
What does bootstrap stability show?
Bootstrap resampling repeats threshold selection on sampled datasets. A narrow range suggests a stable cutoff. Wide variation indicates greater selection uncertainty.
Should thresholds be monitored after deployment?
Yes, prevalence and score distributions can change later. Costs and operational capacity may also shift. Regular monitoring keeps decisions aligned with reality.