Analyses is the catalogue of ready-made questions you can run on your book. Each analysis uses a defined order of steps, tests the evidence, and checks known traps, so you can see what the data supports and how certain it is. The method behind an analysis is also called a playbook.
- What an analysis does
- Choose an analysis
- Read the result
- Treat watch-outs as limits
- Know what the run can change
- Check the data before trusting the answer
What an analysis does
Each analysis has four parts:
- The question it answers
- The data it needs
- The Hypotheses it tests
- The Watch-outs it checks
An analysis can also produce charts or tables. Some analyses build a decision flow, such as a scorecard decision flow, price grid, collection queue, or work queue.
The catalogue addresses questions about your portfolio, losses, decisions, scores, models, and emerging risk. It does not decide what you should do. A rejected hypothesis or a fired watch-out is a finding for a person to assess and act on.
The existing Analysis page covers questions about decisions already taken, including backtests and mining. Those runs appear in the same Analyses list as the analyses described here.
Choose an analysis
Open Analyses from the left menu and select New analysis. The dialog lists every available analysis, with ready analyses first, and shows what is missing for the others.
| Group | Analyses |
|---|---|
| Portfolio and losses | Collection curves, Performance drivers, Vintages, Collections, State moves, Riskiest segments, Unit review, Expected loss, Provision, Provision back-test, Portfolio pricing, Fraud investigation, Collection queue, Work queue |
| Decisions | Decision summary, Risk report, Decision backtest, Champion and challenger, Shadow review, Cut-off strategy, Decision flow on records |
| Scores and models | Characteristic strength, Score performance, Model stability, Scorecard from an outcome, Scorecard from decisions, Scorecard from records and events, Score a new tape |
| Watching for risk | Concentration, Connected exposure, Group exposure, Identity rings, Shared identifiers, Repeated identifiers, Leading digits, Ratio over a limit, Company registry watch, Connections watch, Digest |
Each analysis states which data it reads. The catalogue uses customer records from a data intake, the decision archive with its outcomes, or the portfolio warehouse. See Data for analyses for the data feeds and their preparation.
Choose attributes in the inputs page. See Choosing inputs. Start or repeat a run, configure email, and open cases from a finding through Running analyses. The group pages explain what each analysis in a group decides and how to read its answer.
Read the result
The result page leads with a summary such as the number of hypotheses that held and the number of watch-outs that fired. It then shows:
- Hypotheses, with the number behind each verdict
- Fired Watch-outs
- Findings that are new since the previous run
- Charts and tables
- Steps
- Earlier runs
A hypothesis is a statement the analysis tests, such as whether fresher segments collect more, whether a score holds out of sample, or whether a cut-off pays. Its verdict is:
- held when the evidence survives a test that could have failed it
- rejected when the data contradicts it
- unclear when the evidence cannot settle it, such as when two ranges overlap
The number beside the verdict lets you challenge the answer. A verdict without its number does not show how strong the evidence is.
A finding can be New, Still open, or Cleared. Use a rejected hypothesis or fired watch-out as a reason for a person to investigate or act on. The decision playbook explains how analyses about decisions present their results.
Steps is a collapsed evidence view. It lists each step the run took and what that step returned. It shows how the answer was reached. It does not provide advice or actions.
Treat watch-outs as limits
A watch-out turns a known trap into a check. Examples include:
- A half-finished last period that makes a recent cohort look weak
- A pricing segment that is also scored, so the effect is counted twice
- A random holdout that leaks a group effect
- Too few outcomes for each band to be more than directional
- Information value so high that the score was probably built from the outcome
- Payments that match no record
A watch-out can stay quiet, warn, or block the run. A warning lets the run finish and puts the caveat beside the result. A block stops the run before the next step because anything built on top of it would be misleading. The run then shows Blocked and explains why.
Directional means the sample is too small for the number to be firm. Treat a directional result as an indication that needs further evidence.
Know what the run can change
Nothing an analysis does changes a deployed release or live traffic. When an analysis builds something, it builds a decision flow and releases it under the name of the person who started the run. That person's rights apply, and the result provides a link to open the decision flow. Nothing goes to an endpoint on its own.
Every number on the page comes from the analysis itself. It does not come from a language model. This gives you a direct basis for checking the numbers against the hypotheses, watch-outs, and steps.
Check the data before trusting the answer
An analysis is only as good as the mapping of its data. A wrong outcome column produces a confident wrong answer. The portal checks files and asks you to confirm the mapping before the data becomes records, and several analyses warn when a result looks too good to be true.
A data intake is a set of your files, including xlsx, csv, JSON lines, or zipped files, or a feed read from the decision archive. The portal reads the data, checks it, then turns it into records. See Data for analyses for the full data-intake process.