Define volatility before measuring it
Volatility means how much a currency exchange rate tends to vary over time. In practical terms, it is usually expressed from changes in price rather than the price level itself. For a USD/CNH quotation, you measure volatility by converting observed exchange rates into returns (for example, how much the rate changes from one timestamp to the next), and then summarizing how variable those returns are.
A key point is to separate stable mechanics from variable conditions:
- Mechanics: the math that turns a time series into a volatility statistic.
- Conditions: what data you use, how often you sample it, and how transaction costs or execution constraints might affect what you experience.
Choose a measurement method (and make the assumptions explicit)
Common volatility measurements share the idea of using returns, but they differ in how they define returns and aggregate them.
1) Historical (sample) standard deviation of returns
- Assumption: you have a sequence of exchange rates sampled at a fixed interval (for example, every hour).
- Return definition example: simple return r_t = (P_t − P_{t−1}) / P_{t−1}, or log return r_t = ln(P_t / P_{t−1}).
- Volatility estimate: the standard deviation of r_t over a chosen window, such as the last N observations.
If you present a number, specify the interval and window length because volatility is not a single universal constant. A 1-week estimate and a 1-day estimate can differ materially even if the market is unchanged.
2) Rolling volatility (time-varying estimate) Markets often change regimes. A rolling approach measures volatility repeatedly: compute the standard deviation of returns inside a moving window (e.g., last N hours) and slide it forward.
- Mechanically, this is the same standard deviation concept, but the output becomes a time series of volatility.
3) Realized volatility from high-frequency data With more frequent observations, you can estimate how variation accumulates within a day. The exact formula depends on the return definition and sampling scheme.
- Assumption: you can sample prices consistently.
- Limitation: higher frequency can increase the impact of microstructure noise (small-scale artifacts in quoted data).
4) Exponentially weighted moving estimates (decay) Another choice is to weight recent returns more heavily than older ones. This can make the estimate respond faster to changing conditions.
- Assumption: “recent matters more” is an intentional modeling choice.
- Limitation: the result can become sensitive to the chosen decay rate.
Worked example using general formulas (no live prices assumed)
Suppose you have exchange-rate observations P_0, P_1, …, P_N at equal time steps. Choose a return definition, such as log returns:
- r_t = ln(P_t / P_{t−1}) for t = 1…N
Then estimate volatility as the sample standard deviation of {r_1,…,r_N}:
- s = sqrt( (1/(N−1)) * Σ_{t=1..N} (r_t − mean(r))^2 )
What you must state up front:
- Sampling frequency: are P_t hourly, daily, or something else?
- Window size: what N did you use?
- Return type: log vs simple.
Material impact of these choices:
- If you change the sampling frequency, the return sequence changes and so does the standard deviation.
- If you switch from simple to log returns, the numerical value may shift (especially when moves are large).
In a realistic scenario, you might compare two windows (e.g., 30 days vs 90 days). A longer window typically smooths short-term spikes, while a shorter window reacts more quickly.
Limitations, failure modes, and what can go wrong
Even if the math is correct, volatility measurement can mislead if assumptions do not match reality.
1) Data and sampling problems
- Missing observations or irregular timestamps break the “equal interval” assumption.
- Outliers (rare jumps due to data errors or unusual prints) can inflate standard deviation.
2) Non-trading and market structure effects FX quoting and liquidity conditions vary through time. If the sampling captures periods with different market depth, the realized variation may reflect trading conditions rather than underlying exchange-rate movement.
3) Cost and execution effects are not included automatically Volatility calculated from mid-market rates does not automatically capture the effect of spreads, slippage, or order execution. If you later try to relate volatility to trading experience, the mismatch can be significant.