VaR Exceptions, Clustering, and a High-Yield Request
A $1.2 billion multi-asset portfolio uses a one-day 99% historical VaR model. Over the last 250 trading days there were 5 exceptions, meaning losses exceeded VaR on five days. The model uses two years of daily data and aggregates positions by asset class. The portfolio has $180 million in high-yield bonds, $90 million in emerging-market equity, and $40 million in a bank-loan ETF that traded at a 4% discount to NAV during the last stress episode. The risk committee is concerned because the exceptions clustered in two weeks during a rates sell-off. The CRO proposes doubling the VaR confidence level to 99.9% and extending the lookback to five years. The CIO wants to keep the 99% model and add stress tests. The portfolio’s limit is a 10% one-month 99% VaR; current one-month VaR is 8.2%.
1. Evaluate the backtesting result. Is five exceptions in 250 days at 99% a model failure, and what does the clustering imply?
Read the worked answer
At 99% over 250 days, the expected number of exceptions is 2.5. Five is twice the expectation. Under a binomial or Poisson approximation, five or more exceptions occur with roughly 10–11% probability even if the model is correct, so five alone is not a decisive rejection at conventional confidence levels. But the clustering matters more than the count. Historical VaR assumes exceptions are independent; two clusters in a rates sell-off suggest volatility clustering and regime dependence. A Christoffersen independence test would likely reject. The model also uses two years of daily data and aggregates by asset class, which may miss within-sector concentration and liquidity risk. The correct response is not simply to declare the model failed, but to treat it as a warning: review data quality, recalibrate volatility, add conditional volatility or filtered historical simulation, and run stress tests. A model that produces exceptions only in one regime is not capturing the current risk environment.
2. Assess the CRO’s proposal to raise confidence to 99.9% and extend the lookback to five years. Will that improve risk management?
Read the worked answer
Raising the confidence level to 99.9% reduces the expected exceptions to 0.25 per 250 days, which makes backtesting almost impossible; a model that cannot be validated is not safer. Extending the lookback to five years adds older, possibly calmer data and may understate current volatility, while also mixing different market regimes. The proposal treats the symptom rather than the cause. Better alternatives are filtered historical simulation or a GARCH-type volatility overlay, stressed VaR using a crisis window, liquidity-adjusted VaR with longer horizons for high-yield and bank-loan exposures, and scenario analysis for rates and credit spreads. The committee should also decompose VaR by asset class and position. A higher confidence number may look conservative, but it can obscure the fact that the model still assumes the past is a good guide to the future.
3. The portfolio’s one-month 99% VaR is 8.2% against a 10% limit. The CIO wants to add $60 million of high-yield. What checks and limits would you impose?
Read the worked answer
The current VaR is $98.4 million against a $120 million limit, leaving only $21.6 million of headroom. Adding $60 million of high-yield raises high-yield exposure from 15% to 20% of the portfolio and may use much of that headroom once correlations are considered. VaR alone is insufficient. Require incremental and component VaR, not just the total, and stress a 200 bp spread widening plus a 10% price decline in high-yield; that alone is a $6 million loss on the new position, before correlation effects. Include the bank-loan ETF’s 4% NAV discount and liquidity risk in the stress. Set a high-yield concentration limit, a stress-loss limit, and a liquidity limit. Approve only if the incremental VaR and stressed loss remain inside limits, and if the CIO agrees to a pre-defined de-risking trigger. If the addition breaches the 10% VaR limit under stress, reduce the size or offset it with hedges.