Evidence of non-stationary bias in scaling by square root of time: Implications for Value-at-Risk

In this paper, we show that scaled conditional volatilities obtained by the square root formula applied to i.i.d residuals from a sample of Canadian stock market data for various time horizons and error distributions, typically underestimate the true conditional volatility; consistently have a higher standard deviation and exhibit non-stationary kurtosis. Furthermore, the bias produced by volatility scaling is non-stationary in mean and standard deviation and its magnitude is likely influenced by monetary policy regime shifts. Moreover, while VaR is risk-coherence for elliptical distributions, this bias remains even for this class of distributions.