Financial risk models that compute Value-at-Risk (VaR) often assume returns are independent or scale simply with time. Real markets exhibit volatility clustering, periods when large moves follow large moves and small moves follow small moves, a phenomenon formalized in the ARCH family of models by Robert F. Engle New York University Stern School of Business and extended by Tim Bollerslev at Duke University. This persistence produces conditional heteroskedasticity and heavier tails in aggregated returns, which directly affects long-term VaR.
Mechanism of bias
When volatility is persistent, short-term shocks do not wash out quickly; instead they propagate into future periods. Conventional long-horizon VaR that scales one-day risk by the square root of time implicitly assumes independent increments. That assumption fails under clustering, so scaling underestimates the probability of extreme cumulative losses. In statistical terms, persistence increases the variance of multi-day returns beyond the naive t-times-one-day variance, and it alters tail behavior. J.P. Morgan’s RiskMetrics historically highlighted the dangers of simple scaling and promoted exponentially weighted volatility estimates to capture time-varying risk, but even such methods can misstate long-run dependence if the underlying process has strong long memory.
Practical corrections and implications
Bias in long-term VaR has concrete consequences for capital allocation, risk limits, and stress testing. Regulators and standard setters such as the Basel Committee on Banking Supervision emphasize model risk and the need for robust backtesting; underestimating VaR can lead to insufficient capital buffers, mispriced risk transfer, and systemic vulnerability during crises. Human and cultural factors amplify these effects: markets in emerging economies, where liquidity is shallow and information flow is uneven, often show stronger clustering, while political instability or environmental shocks can create localized persistence that propagates across borders through investor behavior.
To mitigate bias, practitioners use conditional volatility models like GARCH, long-memory specifications, scenario-based Monte Carlo, and historical simulations that preserve temporal dependence. Combining statistically grounded models from Engle and Bollerslev with rigorous governance, localized market knowledge, and stress scenarios improves long-horizon VaR reliability, acknowledging that no model perfectly captures tail dynamics but some approaches substantially reduce systematic underestimation.