Home / Use cases

Valcision use cases

Fraud risk modeling

Develop transaction-scoring workflows, compare models, and align review thresholds with investigation capacity. Keep the distinction between a suspicious flag and confirmed fraud clear.

Turn transaction patterns into reviewable evidence.

Fraud risk modeling combines data preparation, scoring, and evaluation. Valcision supports labeled classification and anomaly detection so teams can examine suspicious patterns and understand the tradeoff between finding fraud and generating false alerts.

  • Which transactions rank highest for investigation?
  • How many labeled frauds does a model find in a later evaluation period?
  • How does precision change as the review threshold changes?
  • Can the proposed alert volume fit the team’s review capacity?

Prepare the data before selecting a model.

Transaction features

Bring relevant numeric transaction attributes and observation times. Choose features that were available when the transaction would have been scored.

Labels and evaluation periods

For supervised development, supply a fraud label and define which value means fraud. Reserve later observations for evaluation and account for delays in label availability.

Exclude the target label from predictors. Review missing values, sampling choices, and repeated customer or account records across development and evaluation periods. A sample with a higher fraud share than the population can produce misleading operational expectations.

Explore the demo workflow

This flowchart is rendered from a saved Valcision application demo.

Import synthetic transactions, create seeded development, validation and test partitions, and fit the Fraud Detection model on development data. The report receives held-out test scores.
Import synthetic transactions, create seeded development, validation and test partitions, and fit the Fraud Detection model on development data. The report receives held-out test scores.

Illustrative workflow using synthetic demo inputs. Click the chart to view it at full size.

Connect model choice to the review decision.

  1. Prepare features. Import transactions, select predictors, and define the label and chronological evaluation split. Fit preprocessing on training data.
  2. Develop candidate models. Use balanced logistic regression with labeled data, or Isolation Forest to explore anomalies when labels are unavailable or as a comparison.
  3. Set a review threshold. Use the configured review fraction to establish a threshold from training scores. Evaluate the actual alert rate on later data.
  4. Measure the tradeoff. With labels, review average precision, alert precision, fraud recall, detected frauds, missed frauds, and false alerts.
  5. Inspect scored records. Examine scores, ranks, suspicious flags, and training or holdout membership alongside the original transaction fields.
  6. Refine and report. Compare candidates and thresholds with review capacity, false-alert costs, and missed-fraud costs in view.

Illustrative use case

All figures below are fictional and deliberately simplified. They illustrate a workflow and calculation; they are not customer results, fitted model outputs, or forecasts.

The evaluation dataset

A fraud analytics team evaluates scored transactions on a later, labeled holdout period containing 100,000 transactions. Of these, 200 are confirmed frauds and 99,800 are not, a fraud prevalence of 0.2%.

The team wants to compare two review thresholds against investigation capacity. The counts below are assumed evaluation results, not performance measured from a trained model.

1. Prepare a development and evaluation split

The team selects features available at transaction time, excludes the fraud label from predictors, and reserves later transactions for evaluation. It checks delayed labels and customer or account overlap between periods. Preprocessing is fitted using development data.

2. Develop and score candidate models

The team develops a labeled model such as balanced logistic regression, with Isolation Forest as a possible anomaly-detection comparison. It retains transaction scores, ranks, and evaluation membership so each flagged record can be reviewed.

3. Compare review thresholds on the same holdout

Evaluation measureWider queueNarrower queue
Transactions flagged1,000500
Confirmed frauds found (true positives)160130
Non-fraud alerts (false positives)840370
Confirmed frauds missed (false negatives)4070
Precision16%26%
Recall80%65%
Review rate1.0%0.5%

Follow the calculations

Precision = confirmed frauds found ÷ transactions flagged. Wider queue: 160 ÷ 1,000 = 16%. Narrower queue: 130 ÷ 500 = 26%.

Recall = confirmed frauds found ÷ all confirmed frauds. Wider queue: 160 ÷ 200 = 80%. Narrower queue: 130 ÷ 200 = 65%.

The wider queue adds 500 reviews to find 30 additional frauds, while generating 470 additional false alerts. The narrower queue halves review volume but misses 30 more confirmed frauds in this sample.

4. Assess the operational tradeoff

Valcision provides scored records and validation outputs. A financial institution then compares to its thresholds, in addition to reviewing capacity, investigation cost, fraud amounts, and label completeness before taking further evaluation.

Limitations of the example

A suspicious flag requests investigation; it is not confirmation of fraud. These scores are not assumed to be calibrated probabilities, and this example does not imply live transaction blocking.

Valcision provide graphical and analytical tools for fraud detection.

Evaluate ranking and operational tradeoffs.

OutputWhat it tells the team
Scored transaction datasetOriginal records with scores, ranks, review flags, and evaluation membership.
PrecisionThe share of flagged transactions that are labeled fraud.
RecallThe share of labeled frauds found at the selected threshold.
Average precisionA summary of ranking performance across precision-recall thresholds.
False alerts and missed fraudsThe investigation workload and frauds not found in the evaluation sample.

Review these outputs together. Strong ranking performance can still produce many false alerts when fraud is rare.

Common questions

Does a suspicious flag confirm fraud?

No. It identifies a record for review under the configured method and threshold. Confirmation requires the relevant investigation and outcome evidence.

Is the fraud score a calibrated probability?

The supported anomaly and class-weighted scores are not presented as calibrated fraud probabilities. Use them for ranking and evaluate them against the intended review decision.

Can the workflow run without fraud labels?

Isolation Forest can score unlabeled data for anomalies. Without labels, the data cannot establish actual fraud precision or recall.

Does the development workflow block transactions in real time?

This page describes model development, evaluation, and scored outputs. It does not imply a real-time transaction-blocking integration.

Bring your analytical need into view.

Share your data, modeling question, and the decisions your team needs to support.

Discuss your use case