Case 007Market risk limits and VaRCore
A bond-futures basis desk with a 99% VaR of Rs 8 crore loses Rs 14 crore in one day. The VaR model maps cash bonds and futures to the same yield curve. Explain the loss and fix the measurement.
1The situation
Kelvora Markets runs a government bond basis desk. It is long cash government bonds and short the matching bond futures, sized so the two legs have equal DV01 of Rs 1.4 crore per basis point each. The trade earns the small, steady difference between the bonds and the futures and has been profitable for two years.
The desk's reported 99% one-day VaR is Rs 8 crore, almost all from its other positions. Yesterday the desk lost Rs 14 crore: cash bonds cheapened sharply against futures while outright yields barely moved. The bank's VaR model maps both the cash bonds and the futures to a single government yield curve.
2Your task
Why did the VaR not warn of this loss, and how should the measurement be changed?
Quick check
What did the VaR model see in the basis position before the loss?
Worked solution
Try it on paper, then open one step at a time.
30-second answerThe answer to give first
The VaR could not see the loss because the risk that caused it was not in the model. Mapping cash bonds and futures to one curve makes a Rs 1.4 crore DV01 long and short cancel, so the basis has no volatility. The basis widened about 10 basis points, costing Rs 14 crore. Add the basis as its own risk factor, measured on stressed history, and cap gross DV01.
Step 1Why did a hedged position lose Rs 14 crore?
Picture holding a fixed deposit and owing exactly the same amount on a loan at the same rate. Rates move and you do not care. Now suppose the loan reprices and the deposit does not. A hedge protects you only against the risk both legs share; the gap between the legs is a separate risk that stays with you. Kelvora's legs share the level of government yields. They do not share the spread between cash bonds and futures, the basisThe difference between the price of a cash bond and the price implied by the futures contract on it; it moves with funding costs, delivery options and demand for each., and that spread is what moved.
The arithmetic is short. Cash bonds cheapened about 10 basis points relative to futures. The desk is long Rs 1.4 crore per basis point of cash bonds, so 10 basis points of basis widening costs Rs 14 crore, with the futures leg offering nothing back because futures yields did not move with the cash bonds.
Step 2Why could the VaR model not warn of it?
A VaR model only measures the risk factors it has. Mapping both legs to one curve is an assumption that the basis never moves, so the model reports zero risk for exactly the position that carries it. This is not a bad day for a good model; it is a gap in the model, and the loss would not have shown up even with a thousand years of history. Risk teams keep a list of such gaps, called risks not in VaRAn inventory of exposures the VaR model does not capture, each measured separately and covered by limits or capital., for exactly this reason.
Step 3How would you fix the measurement?
Map the cash bonds to their own bond-specific yields and the futures to the futures price, so the spread between them becomes a risk factor with its own history. At a calm-period basis volatility of 1.5 basis points a day, 99% basis VaR is about Rs 4.9 crore; on a stressed window with 4 basis points it is about Rs 13.0 crore, close to the actual loss. Use the stressed figure for limits, because basis trades earn small amounts in calm markets and lose large ones in stress.
| Measure | Basis volatility, bp a day | 99% basis VaR, Rs crore |
|---|---|---|
| One curve for both legs | 0.0 | 0.0 |
| Separate basis factor, calm history | 1.5 | 4.89 |
| Separate basis factor, stressed history | 4.0 | 13.05 |
Add two controls that do not depend on any model: a gross DV01 limit on each leg, so the desk cannot scale the trade just because its net risk looks small, and a basis stress test, such as the widest one-week basis move on record, reported beside the VaR. The limitation to say: even a separate basis factor relies on history, and a basis driven by a funding squeeze can move further than it ever has.
Where candidates lose it
Candidates say the loss was a 1-in-100 day that VaR is allowed to miss. That misreads it: the model could not have shown the risk on any day, because the basis was not one of its inputs.
The second miss is fixing it by raising the VaR confidence level to 99.9%. A higher percentile of zero risk is still zero; the fix is a new risk factor, not a new percentile.
What the interviewer asks next
- How would you backtest the new basis factor?
- The desk argues its net DV01 is small, so its limits should stay as they are. How do you respond?
- What market events typically make the cash and futures basis blow out?
Company names and figures are illustrative.
