Internal Model Validation Findings Before a Dividend Decision
An internal-model validation of Aurora General's economic capital model has landed on your desk. The liability-risk module uses a factor-based formula calibrated on 2016–2019 data and excludes the 2020–21 claims shock. The dependency structure assumes a 0.35 correlation between motor third-party reserve risk and property catastrophe risk; the 2020 back-test showed realised co-movement closer to 0.60. Two key parameters are entered manually each quarter in a spreadsheet with no version control, by an analyst who is the only person who understands the code. The model produces a capital requirement of ₹1,850 crore and an indicative solvency ratio of 1.58. The board wants to use the output both to set the reinsurance retention and to pay a ₹120 crore dividend. The chief risk officer has asked for your view within five days.
1. Assess each finding for materiality and for the direction in which it biases the capital number.
Read the worked answer
Rank by direction and size, not by how easy the finding is to fix.
- Excluding 2020–21. Almost certainly understates liability risk. That shock is a real observation, not an outlier to be discarded; excluding it is a calibration choice that flatters the answer. Material if it moves capital by more than roughly 5%.
- Correlation 0.35 versus realised 0.60. This is the most material single input, because motor third-party reserves and property catastrophe are the two largest risks in the portfolio. A higher correlation reduces diversification and raises aggregated capital. Test the effect: if it adds ₹150 crore, the ratio falls from 1.58 towards 1.45 - below the control level.
- Manual spreadsheet entry and key-person dependency. Not a number, but a model-risk finding: the output is not reproducible, there is no audit trail, and a single analyst's absence or error changes the capital number silently. This is a governance failure as much as a technical one.
- Back-test failure. The model did not predict observed co-movement over a short calibration window. It is evidence that the parameters are not stable, not merely that one year was unusual.
Note that the two main biases pull in opposite directions. Test them together, not separately, or the offset will be missed.
2. What interim actions would you take before the model is rebuilt?
Read the worked answer
Do not freeze decision-making while the model is rebuilt - overlay it.
- Apply a stressed correlation. Replace 0.35 with 0.60 in a copy of the model and re-run. Report the revised capital number alongside the original, not instead of it.
- Add a liability-risk margin. Include an explicit allowance for a 2020–21-type event, either as a scenario uplift to the liability charge or as a management overlay in the capital assessment.
- Run a small scenario set as a cross-check: a 10% motor third-party reserve deterioration; a 1-in-100 cyclone; and the two combined. If the ratio stays above 1.50 with a buffer in all three, the dividend case becomes credible.
- Replicate independently. Have a second analyst rebuild the calculation from source data and reconcile to the model output; this addresses both the spreadsheet risk and the key-person risk.
- Control the environment now. Freeze the spreadsheet version, add a change log and dual sign-off for parameter entry, and schedule migration to controlled code.
Document all overlays in a model-limitation memo signed by the chief risk officer and the chief actuary, with a capital range (for example ₹1,750–2,050 crore) rather than a single figure. State explicitly that the range, not the point estimate, is the basis for decisions.
3. What would you tell the board about the dividend and the reinsurance retention, and what governance steps follow?
Read the worked answer
To the board, state plainly: the model is not currently fit for capital decisions without overlays; the point estimate is uncertain within a range; the direction of bias is mixed but the tail is understated; and a ₹120 crore dividend consumes capital that may be needed as buffer.
Decision rule. Approve the dividend only if the ratio remains above 1.50 under the stressed correlation and the full scenario set, with a buffer of, say, 10 points. If not, recommend a reduced or deferred dividend and say what would need to change. For the reinsurance retention, the stress result matters more than the point estimate - set retention so that the retained loss in the 1-in-100 scenario is survivable, not so that the model's central case looks efficient.
Governance steps: report to the Model Risk Committee and Audit Committee; record the validation findings with owners and deadlines; require annual independent validation and formal change control; and adopt a model-risk policy covering documentation, replication and key-person dependency.
If the model underpins the ORSA or regulatory capital, disclose the issue to the regulator and correct any filing that is materially wrong. Signing off a number you know to be biased is a conduct issue, not a commercial one.