The rolling volatility panel on the Distributions tab has a three-button toggle for 30D, 90D and 1Y, and on the capture I am working from the 30-day line reads 69.3% at its latest point, over a window running from 2026-04-11 to 2026-07-11. The Risk and Statistics block on the overview tab of the same page reports an annual volatility of 42.30% over 86 observations spanning 96 days.
Those two figures are not in conflict. They are different windows on the same book, and the gap between them is the entire subject of this piece, because a vol target that resets off the first number and a vol target that resets off the second run at materially different gross exposure. Before you change the multiplier on a live book, you owe the process a decision about which estimator you are trusting and why.
What a reset actually does to the book
Volatility targeting scales positions by the ratio of target to realised. If realised doubles, gross halves. That is the whole mechanism, and it is well behaved when realised volatility is measured accurately and badly behaved when it is not, because the estimator's noise is transmitted directly into turnover.
Assume iid returns for a moment, which is wrong but tractable. The sampling standard error of an estimated standard deviation from n observations is approximately the estimate divided by the square root of twice n. Run that on the three settings the panel offers, on a daily series:
- 30 observations gives a standard error of about 12.9% of the estimate. On a 69.3% reading that is roughly plus or minus 9 volatility points at one sigma.
- 90 observations gives about 7.5%, or roughly 5 points on the same reading.
- 252 observations gives about 4.5%, or roughly 3 points.
So a 30-day estimator that moves from 55% to 69% has not necessarily told you anything. That move is inside a one-sigma band of an unchanged process. Fat tails and volatility clustering make the real band wider than the iid figure, not narrower, so treat 12.9% as a floor on your uncertainty rather than a measurement of it.
The single-bad-week test
The question the three lines are there to answer is whether an elevated short-window reading reflects a level shift or one event still sitting inside the window. There is arithmetic for this and it takes a minute.
Suppose variance is estimated as the mean squared return over n days. One outlier of size k times the typical daily move replaces one ordinary contribution, so the new variance is the old variance times (n minus 1 plus k squared) divided by n. On a 30-day window, a single five-sigma day gives (29 plus 25) over 30, which is 1.8, and the volatility estimate rises by the square root of that, about 34%.
One day. A third of the reading. And it drops out of the estimate entirely on a known date, thirty observations after it occurred.
That gives you a concrete diagnostic rather than a feeling. Find the largest one or two absolute daily returns inside the short window, compute what the estimate would read with them excluded, and mark the calendar date on which they roll out. If the trimmed estimate sits near the 90-day line, you are looking at an event, not a regime, and resetting gross now means you will reset it back on a date you can already predict. That is round-trip turnover you chose to pay.
Reading the three windows as a set
Four configurations, four different decisions.
Short elevated, medium and long flat. An event inside the short window. Run the trim test. If it confirms, either do nothing or apply a partial reset with a documented rationale, and diarise the roll-out date.
Short and medium both elevated, long turning up. A level shift with confirmation from an estimator that has half the noise of the short one. This is the configuration a vol target exists to respond to, and it is also the one where waiting for the long window costs you real money, so the medium line is the one that should carry the decision.
Short below medium and long. Compression, and the dangerous one. A vol target is a leverage instruction, and a falling denominator instructs you to lever up, into precisely the conditions from which volatility tends to expand. Every vol-targeted book needs a gross cap that binds independently of the estimator, and this is the configuration that finds out whether yours does.
All three elevated and roughly equal. The new level is the level. Reset, and reset the target too if the strategy's capacity assumptions were built on the old regime.
The conventions that have to be written down first
Three parameters change the number materially and none of them is visible on the face of a chart.
The annualisation factor is the first. A crypto book trades 365 days and an equity book 252, and the same daily series annualised on the two conventions differs by the square root of 365 over 252, which is a factor of 1.20. A book reading 42.30% on one convention reads about 50.9% on the other, with nothing whatsoever changing in the data. If the risk function and the reporting function have made different choices here, the same portfolio carries two volatilities and two Sharpe ratios, and the argument that follows is unwinnable because both sides are right.
The return frequency is the second. Daily, hourly and weekly returns on the same book do not scale into each other cleanly once there is any autocorrelation, and trend-following books have plenty of it. Pick a frequency for the risk process and use it everywhere.
The estimator itself is the third. A simple rolling window weights every day inside it equally and drops the oldest day with a cliff. An exponentially weighted estimator decays smoothly and has no roll-out date at all, which removes the artefact described above at the cost of never fully forgetting an event. Both are defensible. Only one of them should be in your policy document.
Turning this into a rule that survives a review
Write the reset policy before the next reset, not during it. Four clauses is usually enough.
Use a band, not a point. Reset only when realised volatility leaves a range around target, for example beyond a quarter of the target in either direction, so ordinary estimator noise does not generate turnover.
Separate the trigger from the level. Let the short window fire the trigger, since that is what a fast estimator is for, and let the medium window set the size you reset to, since that is where the noise is halved.
Impose a minimum interval between resets and a cap on the leverage multiplier. Both bind hardest exactly when you least want to obey them, which is what makes them worth having.
Compute the implied annual turnover of the rule before it goes live, in basis points, and put that figure next to the expected risk reduction. A vol target that costs more in execution than it saves in drawdown is a common and entirely avoidable outcome, and the arithmetic that reveals it is available before any money moves.
One last note on sourcing. The rolling panel and the risk block on this page are computed over different windows and one of them exposes its own risk-free field and observation count in its footer. Any figure you carry out of here should be recorded with the panel it came from, the account scope, the date range and the toggle position, because the same metric name on two Blockcircle surfaces need not carry the same value, and reconciling that after the fact is far more work than noting it at the time.