Measuring Volatility Properly: Realized Variance, True Range and Beyond
Posted: Mon Oct 26, 2026 1:39 pm
Volatility is not observed directly, so every measure is an estimate. Several common ones answer slightly different questions, and mixing them up leads to inconsistent parameters.
Realized variance sums squared returns sampled at a fixed interval over a period, such as five-minute returns across a day. Sampling more often captures more information but brings in microstructure noise, which is why intermediate sampling intervals are often preferred. Average true range accounts for gaps between bars and is easy to compute, though it is sensitive to the chosen smoothing length.
Range-based estimators, including the Parkinson and Garman-Klass approaches, use the high and low of each bar and generally extract more information from the same data than close-to-close returns. They carry assumptions about continuous trading and no drift that real data only approximately satisfies.
For scalping, the key decision is the window. A very short window reacts quickly but is noisy, while a long one is stable but late. Comparing a short-window estimate with a long-window one gives a ratio that flags expansion or contraction. Whichever measure is chosen, consistency matters more than sophistication, since stop distances and targets should be expressed in the same units as the measure used to size them.
Realized variance sums squared returns sampled at a fixed interval over a period, such as five-minute returns across a day. Sampling more often captures more information but brings in microstructure noise, which is why intermediate sampling intervals are often preferred. Average true range accounts for gaps between bars and is easy to compute, though it is sensitive to the chosen smoothing length.
Range-based estimators, including the Parkinson and Garman-Klass approaches, use the high and low of each bar and generally extract more information from the same data than close-to-close returns. They carry assumptions about continuous trading and no drift that real data only approximately satisfies.
For scalping, the key decision is the window. A very short window reacts quickly but is noisy, while a long one is stable but late. Comparing a short-window estimate with a long-window one gives a ratio that flags expansion or contraction. Whichever measure is chosen, consistency matters more than sophistication, since stop distances and targets should be expressed in the same units as the measure used to size them.