A playbook is the method behind an analysis: a ready-made way to test a risk question. This page helps you organise those analyses around your department’s routine, from noticing a change to deciding what to do about it. The platform supports the analyst’s part of that work and stops where a risk manager signs.
- The department’s loop
- The data behind the questions
- Questions by role
- Set a useful rhythm
- Read the answer before acting
- Why the platform works this way
The department’s loop
A risk department notices something, investigates it, reports it to management and proposes a change for a person to approve. Analyses support that loop. They do not change a live decision flow, cut-off or limit on their own.
- Watch. Schedule analyses for the department’s standing questions, with rolling dates. Starting and repeating analyses explains how to set that up.
- Notice. When a hypothesis fails or a watch-out fires, the analysis can create a finding. The finding is remembered across runs as New, Still open or Cleared, so your team can focus on what changed.
- Investigate. Some analyses start a follow-up when a check fires. For example, a score drift can lead to a search for the segments that moved. Today, Score performance and Repeated identifiers can do this. Other analyses do not yet. The chain is short, bounded and limited to analyses that exist. The results show these under Follow-ups it started.
- Report. New findings can reach their owners by email with the numbers. Each repeating analysis has a tile on Home, and the Digest analysis gives one view of everything open. The sentences in emails and on the page come from the checks themselves, not from a language model.
- Act. Open case gives a finding an owner, tasks and a history in a case design. A repeating analysis can use Open cases in to choose the case design. See Case design. A change to a decision flow goes through the usual path of drafts, backtests, approval and release. Backtest, shadow review, and champion and challenger analyses provide evidence for that path.
The platform does not yet write a management memo automatically or deliver a proposed change as a ready draft. A person turns a finding into a change.
The data behind the questions
A risk department works with three kinds of data:
| What you hold | Where it comes from |
|---|---|
| Applications and decisions: who applied, what they sent, and what the decision flow did and why | The decision archive, with outcomes joined as they arrive |
| The portfolio: what happens to the book afterwards, including dues, repayments, arrears, collections, restructures, write-offs and recoveries | Records with dated events from your files or a feed, or month ends in the portfolio warehouse |
| The outside world: for company lending, registry status, insolvency history and overdue filings | Company registry watch reads a company registry for every company in the book. Analyses do not read adverse media or insolvency notices today. |
A data intake is a set of your files, such as xlsx, csv, JSON lines or zipped files, or a feed read from the decision archive. The portal reads, checks and maps it into records. On data pages, Data intake has the steps Files, Check, Map, Records and Analyses. The data guide explains data setup, and Production data and outcomes describes the connected sources.
Countries and lenders use different column names, while the concepts stay consistent: exposure, start date, score, predicted probability, loss share, outcome, segment, identifier, amount, product and channel. When you map an intake, each column is bound to one of these concepts. An analysis can then select fields by meaning, such as “the score” or “the exposure”, and run on a schedule without someone answering questions. A new country or lender needs a new mapping, not a new analysis. That is why the same analysis can run on a Polish and a South African book.
The data stays in your own analytics database. Analyses can use only data you have connected, and their results depend on the quality of your mapping.
Questions by role
Choose an analysis for the decision or risk question you need to answer. The New analysis dialog groups methods under Portfolio and losses, Decisions, Scores and models and Watching for risk. The overview of analyses explains the methods. Optional inputs are under More settings in the start dialog. Use the pages for portfolio and losses, decisions, scores and models and watching for risk for the questions in each group. The inputs guide explains how to choose inputs.
| Role | Questions and analyses |
|---|---|
| Head of risk | Compare expected loss with appetite. Use Concentration and Group exposure to examine connected borrowers. Use Riskiest segments to see which segments stand out. Use Vintages to compare whether newer cohorts are worse than older cohorts at the same age. |
| Credit risk analyst | Use Score performance and Characteristic strength to assess scores and their inputs. Use Cut-off strategy to examine where to cut and what that cut pays. Build a Scorecard from outcomes or from decisions. Use Decision backtest before releasing a change, and Decision summary for the daily approval rate. |
| Model validation | Use Score performance to check stability, calibration or a second score. Use Model stability to examine inputs that drift or go missing. Use Shadow review and Champion and challenger to gather evidence before a change. |
| Fraud lead | Use Repeated identifiers, Shared identifiers, Identity rings, Fraud investigation, Leading digits and Unit review for stores and agents to find patterns. A rule based on a pattern goes into the decision flow. The analysis finds the pattern. Real-time blocking belongs in the decision flow, not in the analysis. |
| Collections | Use Collections to examine cure and promises kept. Use State moves to examine roll forward and cure. Use Collection curves to examine collections over time. Use Collection queue and Work queue to organise calls. |
| Finance and provisioning | Use Provision against the previous month end. Use Provision back-test before changing coefficients. Use Expected loss under stress to examine stressed loss. |
| Compliance and conduct | Use Ratio over a limit for affordability or loan-to-value limits. Use attribute roles to keep restricted characteristics out of every segment and score. |
Set a useful rhythm
Treat this as a starting point, not a rule. The right cadence depends on how quickly the signal changes and when the outcomes become available.
| Cadence | Analyses to consider |
|---|---|
| Daily | Decision summary and Repeated identifiers |
| Weekly | Shadow review while a change is in test, Concentration and the Digest |
| Monthly | Provision, Collections, State moves, Vintages and Ratio over a limit |
| Quarterly, or when outcomes have matured | Score performance, Model stability, Cut-off strategy, Scorecard refit and Provision back-test |
Fraud signals can decay in days. Loan outcomes can take months to appear, so last week’s applications may be too recent to judge a score. For outcome-based analyses, use a cohort old enough for the outcome to have happened.
Read the answer before acting
Start with the verdict and its number, rather than the chart. A rejected hypothesis with a number is an input to a decision. Unclear also gives useful information: the data cannot settle the point, so do not act on that result.
A hypothesis is a statement the analysis tests. It has a verdict of held, rejected or unclear, with the number behind it. A watch-out is a known trap the analysis checks for. It can stay quiet, warn or block the run. Read Watch-outs before relying on the headline, because a fired watch-out tells you how far to trust the number beside it. A result marked Directional means the sample is too small for the number to be firm. It needs more outcomes before it moves policy.
Be cautious when a result looks too good. Very high separation can mean the outcome leaked into the score. Compare like with like, such as groups of the same age, the same product or the same exposure basis.
Findings help your team track what a run has raised. 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, not advice or a set of actions. A finding is not a decision. The analysis tells you what the data says and how sure it is. A person decides what to do. Anything that changes live decisions is tested on history, then in shadow, then released through the normal approval path.
Why the platform works this way
A department’s expertise sits in the order of its steps, its tests and the traps it knows to check. Each analysis carries those checks so the method can be repeated. Remembered findings let your team manage by exception. People stay in charge of changes because a wrong model can cost more than a delay. A shared vocabulary lets a method learned on one book work on another after that book is mapped.
There is no trigger on an event. The AI assistant cannot start an analysis yet. Company registry monitoring does not cover adverse media or insolvency notices.