A credit scorecard is a snapshot. It captures how borrowers looked and behaved on the day it was built, while customers, products, policies and the economy keep moving. When the live population drifts far enough from the development sample, scores stop meaning what they used to. The population stability index (PSI) and the characteristic stability index (CSI) are the two standard checks for catching that drift early. This guide explains how both work, how to read them, and how to build them into routine scorecard monitoring.
Drift rarely has a single cause. The most common sources are:
Telling these apart is the real job of monitoring, and it starts with a measure that is easy to calculate.
The population stability index (PSI) is a statistic that measures how much a score or variable distribution has shifted between two samples, usually the model development population and a recent population. A higher PSI means a larger shift, which suggests the scorecard may now be scoring customers unlike those it was built on.
Split the score range into bins, using bin edges fixed at development. For each bin, compare the share of the recent (actual) population with the share of the reference (expected) population, then sum across bins:
PSI = Σ (Actual% − Expected%) × ln(Actual% / Expected%)
Each bin’s term is never negative, because the difference and the logarithm always share the same sign. The total is zero when the two distributions match and grows as they diverge.
Take a five-band score distribution. The expected column is the development sample; the actual column is last month’s applications. In the formula, use proportions (0.14, not 14), or the result is inflated 100-fold.
| Score band | Expected % | Actual % | Actual − Expected | ln(Actual / Expected) | PSI term |
| Below 580 | 10% | 14% | 0.04 | 0.3365 | 0.0135 |
| 580–619 | 15% | 18% | 0.03 | 0.1823 | 0.0055 |
| 620–659 | 25% | 26% | 0.01 | 0.0392 | 0.0004 |
| 660–699 | 30% | 26% | −0.04 | −0.1431 | 0.0057 |
| 700 and above | 20% | 16% | −0.04 | −0.2231 | 0.0089 |
| Total | 100% | 100% | 0.0340 |
The total PSI is about 0.034, which sits in the stable range. The population has shifted slightly towards riskier score bands, but not enough to act on by itself. Notice that the two lowest bands gain share while the top two lose it. A pattern like that is worth watching over several months, because a single stable reading can hide a slow trend.
Most lenders read PSI against three bands. These are industry rules of thumb, not statistical laws, so treat them as triggers for review rather than automatic verdicts.
| PSI value | Interpretation | Typical response |
| Below 0.10 | Stable: no significant shift | Continue routine monitoring |
| 0.10 to 0.25 | Moderate shift | Investigate drivers with CSI and monitor more closely |
| Above 0.25 | Significant shift | Escalate, review performance, consider recalibration or redevelopment |
Your own PSI thresholds may differ. Bin count, sample size and portfolio type all affect the value, and many institutions set limits in their model risk policy.
The characteristic stability index (CSI) applies the same logic to a single scorecard variable, such as age band, income band or months on book. It compares how that variable’s attributes are distributed today with the development sample.
PSI tells you the population has moved. CSI tells you which characteristics moved it. That is why the two are used together.
| Aspect | PSI | CSI |
| What it measures | Shift in the final score distribution | Shift in one characteristic’s distribution |
| Level | Whole scorecard | Individual variable |
| Bins | Score bands | Attribute bins of each variable |
| Key question | Has the overall population changed? | Which variable is driving the change? |
| Typical use | Headline early-warning alert | Diagnosis and root-cause analysis |
| Formula | Σ (Actual% − Expected%) × ln(Actual% / Expected%) | Same formula, applied per characteristic |
Suppose applicants under 25 rise from 12% to 20% of the population, the 25 to 40 group moves from 48% to 46%, and the over-40 group falls from 40% to 34%. Applying the formula to each age band gives terms of about 0.041, 0.001 and 0.010, so the CSI for age is roughly 0.05. Almost all of it comes from the under-25 band, which points straight to the source of the shift.
One caution: some practitioners weight the CSI by the points assigned to each attribute, which expresses drift as an average change in score points. Pick one definition, document it, and apply it consistently.
PSI and CSI only protect you if they run on a schedule with clear rules. A practical set-up for scorecard monitoring:
Where volumes allow, run the same reports by segment, such as product, channel or region. A stable overall PSI can hide a large shift inside one segment that is offset by another. Keep a short written log of each review, including the PSI values, the explanation for any amber or red reading, and the decision taken. This trail is what validators and auditors will ask to see.
PSI and CSI detect data drift: the inputs or scores look different from before. They cannot detect concept drift, where the link between characteristics and default changes while the distributions stay the same. Model drift in practice is often a mix of both, which is why stability reports should always sit next to performance reports.
A PSI below 0.10 is generally read as stable, 0.10 to 0.25 as a moderate shift worth investigating, and above 0.25 as a significant shift. These are widely used rules of thumb rather than statistical laws, so check your own institution’s model monitoring policy for the limits that apply to your portfolio.
PSI measures how the overall score distribution has shifted, while CSI measures the shift in a single scorecard variable. PSI tells you that the population has changed; CSI helps you find which characteristics caused it. Analysts usually track PSI as the headline alert and use CSI for diagnosis.
Monthly is common for high-volume retail portfolios, and quarterly can be enough for smaller or slower-moving ones. The right frequency depends on application volume, how quickly the portfolio changes, and your model risk policy. Calculate it more often after policy changes, new acquisition channels or major economic shocks.
Yes. PSI only shows whether the population looks the same, not whether the score still ranks risk correctly. The relationship between characteristics and default can change even when distributions are stable. That is why PSI should be reviewed alongside Gini or KS, bad rates by score band and calibration checks.
Understanding the difference between an application vs behavioural scorecard is the first step. Building one, from binning and WOE to validation and cut-off strategy, is where the skill really develops. If you want to learn this end to end, with real-world case studies and tools used in banks, explore our Credit Risk Scorecard course.
Explore Dexlab Analytics’ Credit Risk Modeling certification program to build PD, LGD, and EAD models from scratch, work through IFRS 9 ECL frameworks, and learn model validation techniques used by practicing risk teams.
.
Credit and Market Risk, Credit Risk, credit risk analysis, Credit Risk Analytics And Modeling, credit risk analytics training, Credit risk certification India, Credit risk modeling course India, Credit Risk Modelling, Credit Risk Modelling Using SAS, Credit scorecard modeling, Expected credit loss ECL, IFRS 9 modeling course India, Logistic regression scorecard, PD LGD EAD modeling, Python credit risk modeling
Comments are closed here.