Watching for risk

This playbook provides analyses for fraud, identity, credit and portfolio monitoring. Each watch looks for a pattern that should not exist, then a repeating run tells you when that pattern changes. Use the result to decide where a person should investigate and which cases need handling.

Choose a watch

The Watching for risk group covers fraud and identity signals, concentration and connections, limits, and a single view of active findings. Choose inputs that match the records and links you can read. A watch is only as good as the identifiers it reads.

Connected analyses depend on a network mapped in the data intake. See Data for analyses for the data that watches read. For starting, repeating, sending findings and opening cases, see Running an analysis.

Fraud and identity

AnalysisWhat it decidesInputsWhat to read
Repeated identifiersWhether the same phone, device, address or national ID returns within a time window, and which repeat threshold best catches the outcome.The identifiers, time field, window in days and number of repeats to flagFor each identifier and for any identifier, how many records a rule would stop, how many known bad outcomes it would catch, its precision and its lift. The watch-out appears when a rule flags more than the set share of all records, because that would swamp reviewers.
Shared identifiersWhether bad outcomes cluster among records joined by shared identifier values.The identifiers used to join recordsHow many clusters contain a known bad outcome, and how many records sit within one, two or three steps of one. Values shared by very many records are left out because they link everyone, and the result reports them as hub values left out.
Identity ringsWhether rings in a mapped network contain several bad outcomes.Minimum ring size and number of bad outcomesThe size of each ring and how many known bad outcomes it holds. Watch-outs identify rings found and rings holding more than one known bad outcome.
Leading digitsWhether the first digits of an amount follow Benford's law overall and by unit. Natural money amounts follow a known pattern, while invented amounts tend not to.Number of leading digits and the largest tolerated deviationThe unit that strays furthest. This points to where to look. It proves nothing alone.

For repeated identifiers, a high catch rate is useful only with the rule's precision and lift. A rule that catches many outcomes but flags too many records creates a review burden, so use the watch-out with the result.

Shared identifiers and identity rings answer different questions. Shared identifiers show how records connect through common values. Identity rings show the size and bad outcomes of rings already found in the mapped network. Treat hub values as a limit on what shared-identifier links can tell you.

Concentration and connection

AnalysisWhat it decidesInputsWhat to read
ConcentrationHow much of the book sits with the largest groups of connected records, overall and by segment.Network, exposure, top N, the share limit for one group and the maximum concentration indexGroups over the share limit and concentration indexes over the limit.
Group exposureWhich entities stand behind several records along directed links, how much exposure sits behind each, and where link weights give an entity a majority.The directed links and their weightsEntities behind more than one record, majority over several records and entities with a large share of exposure. Links into entities whose weights could not be valued count as no majority.
Connected exposureHow much of the book is within a few connections of an entity with a flag, such as a company in liquidation, and which groups are affected.The field used to flag an entity, the values used for that flag, number of hops and exposureWhether anything is flagged, which groups are reached and how much exposure is reached.
Connections watchWhether connections reported by a data source have appeared or ended, and how those changes regroup the book.The decision flow that reads the data source, the connections to read, a budget of reads per run and optional refresh daysNew connections, connections that ended, groups where one connection ended and whether nothing was read. The analysis reads each record, or each not-yet-read node of the graph, keeps the connections it reports and ends connections that are gone.
Company registry watchWhether companies in the book have a failure status, insolvency history or overdue filings in the registry.The companies in the book and the registry they are read fromCompanies flagged, numbers not found and companies that could not be read. The analysis blocks when no record could be checked.

Use Concentration when the concern is the size of the largest connected groups. Use Group exposure when directed links identify entities behind several records. Use Connected exposure when a flag on one entity may affect nearby groups.

Connections watch depends on the source returning current connections. Its read budget limits how many reads a run makes. A result that says nothing was read gives you no new connection evidence for that run.

Limits and the single view

AnalysisWhat it decidesInputsWhat to read
Ratio over a limitWhether the share of records whose ratio of two fields is over or under a limit is high overall or within a segment.Numerator, denominator, over or under, maximum share and minimum records per segmentThe distribution and the worst segment. The ratio can represent loan to value, debt to income, coverage or another ratio.
DigestWhich findings remain open across every active watch.The active watchesOne place to review open findings across the monitoring set.

Use Ratio over a limit when the risk is expressed as a relationship between two fields. The minimum records per segment keeps the segment comparison tied to the records included in the analysis. The worst segment shows where the limit is most concentrated.

Read the result

A result separates the claim under test from the traps that can weaken it.

  • Hypotheses contains statements the analysis tests. Each has a verdict of held, rejected or unclear, with the number behind the verdict.
  • Watch-outs contains known traps. A watch-out stays quiet, warns or blocks the run.
  • Findings contains the issues that need attention. A finding can be New, Still open or Cleared.
  • Steps is a collapsed detail view of each step the run took and what it returned. It is evidence of how the answer was reached. It is not advice and holds no actions.
  • Earlier runs provides the place to review prior runs.

A result marked Directional has a sample that is too small for the number to be firm. Treat that number as an indication rather than a firm conclusion.

Leading-digit analysis is directional in its meaning even when it produces a deviation: it points to where to look and proves nothing alone. For other watches, use the affected records, groups or entities named by the finding to decide what requires review.

Put monitoring into the decision flow

Put the analyses you would otherwise check by hand on a schedule. Use Every day for fraud and identity watches. Use Every week or Every month for concentration and ratio watches. Send findings to the person who owns them, and open cases in the case design that handles them. See Running an analysis for repeating runs, findings and cases.