The Portfolio and losses group contains analyses for understanding how a book performs, valuing it, managing provisions and directing collections. Each analysis follows a playbook, the method behind the analysis, and ends with a decision such as changing a price, provision or collection order. Use Data for analyses to prepare records and events, then use Running analyses to start an analysis and manage its results.
- How to read the answer
- Reading how groups perform
- Valuing and pricing a book
- Provisioning and month end
- Directing collection work
How to read the answer
An analysis states the question it tests as a hypothesis. The result gives the number behind the hypothesis and marks it held, rejected or unclear. A watch-out is a known trap that the analysis checks. It stays quiet, warns you or blocks the run.
A result can also be Directional. This means the sample is too small for the number to be firm. Treat the result as a prompt for review rather than as a basis for a change. Analyses that rank groups state how many records a group needs.
Analyses that use events use complete periods only. When an analysis compares groups, choose groups that have the same relevant age or observation period where the analysis calls for it.
Reading how groups perform
These analyses use records and events to show how outcomes develop, compare groups and identify where to investigate.
Collection curves
Question: How much has each group paid back, month by month since it started?
Choose the grouping for the curve, such as product, vintage or channel. The analysis counts each curve from the start of exposure and uses only complete periods.
The curve shows the shape of repayment rather than one summary number. Use its shape to see how quickly cash arrives and whether repayment has a long tail. Use the answer to review collection expectations or change how you value a group.
Performance drivers
Question: Which groups pay back better or worse than expected, and over what time?
The analysis compares each group with what groups of the same kind would be expected to do. It holds constant the factors you name, usually age or product, and gives a group credit only to the extent that its size supports the conclusion.
Compare groups at the same age. Groups observed for very different lengths of time can make older groups look better. A small book is Directional.
Use the result to investigate the groups that perform differently from their comparable peers. Change a pricing or collection assumption only when the difference is supported by enough records at the same age.
Riskiest segments
Question: Which segments have an outcome rate significantly above the rest?
The analysis tests all segments at once. This prevents testing many segments from manufacturing a culprit. You can also test within another segment.
Choose the minimum records per segment, the confidence level and how many times worse than the rest a segment must be to count. The result gives the segments that stand out and their rates.
Use the result to review a segment, change a cut-off or investigate why its outcome rate is higher. Treat small segments according to the stated record requirement.
Unit review
Question: Which stores, agents or branches need review?
The analysis checks each unit against its own group. A store is compared with its region rather than with the whole book. It also checks for sudden growth between two periods and for amounts above a reference, such as a loan far above the item's price.
Choose the unit, its home group, the date and periods for growth, and the amount and reference to compare.
Use the result to review a store, agent or branch whose outcome rate, growth or amount is unusual.
State moves
Question: How do records move between states from one event to the next?
The analysis shows records that stay, roll forward, cure or exit. You can band numeric states such as days past due.
Use the moves to see whether arrears are building or curing and how one bucket feeds the next. Few moves are Directional, so use a sparse transition as a prompt for more observation rather than as a firm change to policy.
Fraud investigation
Question: Where does the outcome concentrate, and did alerts fire on the records that became fraud?
The analysis combines repeated identifiers, shared identifiers and segments where the outcome concentrates. Restrict it to the part of the book you name. It also checks whether your alerts fired on the records that later turned out to be fraud.
One hypothesis is that the outcome concentrates in a segment. A watch-out applies when alerts fired on less than half of the outcome records.
Use the result to investigate the affected segment and review the alert logic when too few fraud records were alerted.
Valuing and pricing a book
Portfolio pricing
Question: What price grid does the book support by segment and score band?
Choose how records are grouped as bought together, the segment to price by and how payments without a record are grouped. Provide the cost share, target multiple, discount rate and minimum records per segment.
The analysis builds the valuation in a sequence:
- It draws collection curves by segment from the first period onward.
- It places a range around the pooled rate so the price carries its uncertainty and includes a conservative case.
- It borrows a tail of late payments from the longest-observed group because young groups have not finished paying.
- It values the book at the discount rate and target multiple you provide.
- It fits a scorecard with group holdout.
- It creates a grid from segment price and score band.
The hypothesis is that fresher segments collect more and that the ranges separate. A blocking watch-out applies when the history is too short to fit a tail. The analysis cannot produce a supported price in that case.
Use the price grid to re-price the book. Review the segment grouping, assumptions and conservative case before using a price supported by a short history.
Expected loss
Question: What is the expected loss overall, under stress and by segment?
The analysis calculates expected loss as predicted probability of default multiplied by loss share and exposure. Provide the predicted probability, exposure and loss share. Loss share can be a field on each record or one number for all records. Also provide the stress multiple and the loss rate you would tolerate as your appetite.
The result shows expected loss overall, under stress and by segment, against that appetite. Use it to see whether the book or a segment exceeds the loss rate you would tolerate. Change a cut-off or review the affected segment when the result is above appetite.
Provisioning and month end
These analyses use the portfolio warehouse to explain the provision, test whether it has matched realised losses and monitor collections.
Provision
Question: How has the latest published month end changed from the month before?
The analysis reports exposure, provision, coverage rate and the share in stage 2. It also reports provision and coverage by stage.
It checks four reasons a provision number can move without the book getting worse:
- coverage rose by more than the tolerance;
- the stage 2 share rose by more than the tolerance;
- a unit has no coefficient for its state, so it is provisioned at nothing;
- the month end is stale because an earlier month was restated.
Use the result to review the month end and raise a provision when the movement is not explained by the book's performance. Check missing coefficients and restatements before changing the provision.
Provision back-test
Question: Did the loss rates in force a year ago predict what was actually lost?
The analysis takes accounts open a year ago and compares predicted losses with losses actually incurred since then. It reports the comparison overall and by state.
Its watch-outs express the hypotheses: more was lost than predicted, so the provision was too thin; fewer was lost than predicted, so it was too heavy; or fewer than twelve months of history is published.
Provide the tolerance. Use the analysis before changing coefficients and again after the change. Raise the provision when losses were above prediction, reduce or review it when they were below prediction, and treat incomplete history as a limit on the conclusion.
Collections
Question: How are recent past-due spells performing against the previous period?
The analysis examines spells that started in the last 12 months. It reports how many cured, how many were lost and how many promises to pay were kept, compared with the 12 months before.
Provide the tolerance for a fall in cure rate and the lowest acceptable share of promises kept. The watch-outs are a cure rate below the tolerated change and promises kept below the acceptable share.
Use the result to change the collection call order when cure performance or kept promises fall below the limits.
Vintages
Question: Are recent monthly cohorts performing worse than older cohorts at the same age?
For each monthly cohort of new accounts, the analysis reports the rate of each outcome you list. It also shows how far into the account's life accounts first became overdue by the chosen number of days.
Compare cohorts at the same age. Use the result to identify a deterioration in recent vintages and change the relevant cut-off or collection response.
Directing collection work
Collection queue
Question: In what order should collection calls be made?
The analysis uses self-cure days, which is how long debts stay past due before they pay on their own. It uses those days to order calls by debt age, builds the ordering as a decision flow and releases it.
A watch-out applies when almost nobody pays unprompted. Waiting is then pointless.
Use the released decision flow to change the collection call order. Review the self-cure result when the queue gives waiting too much weight.
Work queue
Question: Which queue should receive each day's work, and which case should be opened?
The analysis runs a released decision flow over the day's positions, such as past-due units with one unit per person. It sends each unit to the queue chosen by the decision flow, in priority order, and opens its case.
Its watch-outs identify units that failed to score, cases that could not be opened and units still waiting for a case.
Use the result to review failed scoring, case opening failures and unworked units. Change the decision flow or operational handling when work is sent to the wrong queue or remains without a case.