A bank's LGD model predicts 40% for secured SME loans. A typical default in the workout data recovers 60% after three years, with collection costs at the end. Compute the realised LGD and judge the model.

The realised LGD is about 58.7%, not 40%. Net of Rs 5 crore of costs the bank recovers Rs 55 crore, but only after three years; discounted at 10% that is worth Rs 41.3 crore at default. The model's 40% equals the headline recovery exactly, which suggests it ignores both cost and time. The model fails on this evidence: expected losses are understated by about a third, and it needs recalibrating with an overlay meanwhile.

Lend a friend Rs 1,000, and get Rs 600 back three years later after paying Rs 50 in reminders and bus fares. You did not lose Rs 400. You lost the Rs 50 as well, and the Rs 550 you finally got was worth less than Rs 550 would have been on the day the friend stopped paying. LGD is the economic loss measured at the date of default, so costs come off and the net recovery is discounted back. Rs 55 crore received in three years at 10% is worth Rs 55 divided by 1.1 cubed, Rs 41.3 crore.

The model's 40% is exactly one minus the headline recovery. That match is the clue: the model appears to have been calibrated on nominal recoveries, leaving out cost and time. The size of the error follows. Expected loss is probability of default times LGD times exposure, so an LGD of 58.7% instead of 40% raises expected loss by about 47% for every loan in the segment, and the capital and provisions built on it are short by the same proportion.

One typical case proves the method, not the calibration. Run the calculation across the whole workout sample, including curedA defaulted loan that returns to paying normally, usually with a very small loss. Leaving cures out of the sample overstates LGD. loans, which lower the average, and open cases still in court, which raise it if they are excluded. Check that losses in bad years are represented, because LGD rises when collateral values fall across a downturn. Check the discount rate too: 10% is a policy choice, and a validator reports how the answer moves if it is 8% or 12%.

The recommendation: the model fails validation for this segment. Until it is rebuilt on discounted, cost-inclusive recoveries, apply an overlay that lifts its output toward the realised level, and report the understated expected loss to the model risk committee with its capital impact.