Fraud Probability Calculator

Calculator Inputs

Transaction Amount

Currency units; higher values can raise risk.

Average Amount (30 days)

Used to measure unusual spending vs baseline.

Transactions in Last Hour

Velocity spikes often indicate automation or abuse.

Account Age (days)

Newer accounts tend to carry higher uncertainty.

Device Risk Score (0–100)

Combine signals like emulator, rooting, and reputation.

Merchant Risk Score (0–100)

Category risk, dispute history, and fraud exposure.

Chargeback Rate (percent)

Higher dispute rates typically correlate with fraud.

Geo Distance (km)

Approx distance between user and transaction location.

Transaction Hour (0–23)

Night hours can elevate risk in some contexts.

Country mismatch (billing vs IP)

Mismatch can indicate proxying or stolen credentials.

International transaction

International routing may add friction and risk.

New payee / first-time merchant

First-time relationships can be higher risk.

Reset Download CSV Download PDF

Example Data Table

Amount	Avg 30d	Tx 1h	Age(d)	Device	Merchant	CB%	Dist(km)	Hour	Mismatch	Intl	New Payee	Expected Risk
80	95	1	420	12	18	0.3	15	13	No	No	No	Low
260	120	3	45	35	25	1.2	120	14	No	No	Yes	Medium
1600	140	12	8	82	70	6.5	3200	23	Yes	Yes	Yes	High

These rows are illustrative; real models should be trained and validated on your own labeled data.

Formula Used

This calculator uses a logistic model, commonly used for probability estimation: p = 1 / (1 + e^-z).

The score z is a weighted sum of scaled inputs: z = b0 + Σ (bi · fi). Features are scaled (e.g., log for amount, normalization for risk scores) to keep contributions comparable.

Tip: Replace demo coefficients with values learned from your historical data using logistic regression, then recalibrate probability using proper validation.

How to Use This Calculator

Enter transaction details and contextual risk signals.
Press Estimate Probability to calculate the probability and risk band.
Review the drivers (scaled features) to understand what pushed risk up or down.
Use the guidance to choose: approve, step-up verify, or manual review.
Download CSV or PDF to keep a record of your session history.

Risk Signals and Feature Scaling

Fraud scoring improves when signals share comparable ranges. This calculator converts raw amounts to log10(amount+1) and compares spending to a 30‑day baseline using log10(ratio+1). Device, merchant, and chargeback inputs are normalized to 0–1, preserving their 0–100 meaning. Velocity is scaled per ten transactions, and distance is scaled per 5,000 km to avoid a single kilometer jump dominating the score.

Logistic Probability Model

The model uses a logistic function: p = 1/(1+e^-z). The score z is a weighted sum of scaled features plus an intercept. Positive weights raise risk as the feature increases, while the intercept sets the baseline for “typical” activity. Clamping z within a safe band prevents numeric overflow and keeps results stable for extreme inputs and rare outliers.

Interpreting Probability Bands

Probability is mapped to operational bands: Low (<0.40), Medium (0.40–0.69), and High (≥0.70). Bands support consistent decisioning and help align teams on playbooks. For example, a 0.78 score suggests step‑up verification, while 0.22 supports standard processing. If your fraud base rate is 1%, you may raise thresholds to reduce false positives; if losses are high, lower them to capture more risk.

Evaluation Metrics and Calibration

Use labeled history to validate coefficients and thresholds. Track AUC for ranking quality, then monitor precision, recall, and false positive rate by band. A practical target is keeping manual review volume within capacity while maintaining high capture on confirmed fraud. Recalibrate probabilities with holdout data (or isotonic/Platt scaling) so a predicted 0.60 aligns with observed outcomes across segments.

Operational Use and Monitoring

Deploy scores with guardrails: log inputs, decisions, and outcomes for auditability. Add reason codes by listing the largest feature drivers, such as high device risk plus unusual velocity. Monitor drift in feature distributions, especially device risk and velocity, and watch for sudden shifts in night‑time activity. Review weekly dashboards for band mix changes, chargeback trends, and approval rates. Regularly document threshold changes, and run champion‑challenger tests to compare new coefficients safely. Periodically retrain and retune weights when channels, rules, or fraud tactics shift.

FAQs

1) What model does the calculator use?

Yes. It uses a logistic score z built from scaled inputs, then converts z to a probability with the sigmoid function.

2) How can I make this match my organization’s data?

Replace the demo weights with coefficients trained on your labeled transactions, then validate on a holdout set and recalibrate probabilities before deployment.

3) Can I change the Low, Medium, and High thresholds?

Treat them as operational cutoffs. Move them up to reduce false positives, or down to capture more fraud. Always measure impact on approvals, losses, and review workload.

4) Should I use this score as an automated approval decision?

No. The calculator is educational. Use it to prototype features and explainability, but production decisions should rely on a monitored, validated, and governed model.

5) What is included in CSV and PDF exports?

CSV contains full session history rows for analysis. PDF is a compact snapshot of recent results for sharing. Both exports are generated from your current browser session.

6) How long is my calculation history stored?

History is stored in the server session and clears when the session expires. For persistent storage, write results to a database with proper access controls and retention policies.

Recent Session History

No calculations saved yet. Submit the form to start tracking.