Reading a drawdown, a volatility figure and a correlation
Three risk numbers that are simple to compute and easy to over-read; what each measures, and the assumptions each one hides.
Drawdown, volatility and correlation are the three summary statistics that appear on nearly every risk display. Each compresses a long series into one number, and each discards something specific in doing so. Reading them well is mostly a matter of knowing what was discarded and over what window.
Drawdown measures a path, not a distribution
A drawdown is the percentage decline from a prior peak to a subsequent trough. Current drawdown, shown here as current drawdown, measures the distance from the highest level reached to the present value; maximum drawdown over one year reports the worst such decline within a window; days in drawdown counts how long the asset has been below its peak. The last of the three is the most neglected and often the most informative, because duration is what makes a decline consequential for anyone with a fixed horizon.
Two properties deserve care. First, the arithmetic is asymmetric: a decline of fifty percent requires a subsequent gain of one hundred percent to return to the prior level, and a decline of ninety percent requires nine hundred. This is not a rhetorical point, it is the reason drawdown and volatility give different answers about the same series. Second, drawdown is defined against a peak, so a figure computed since an all-time high, visible as distance from all-time high, is measuring against a single observation that may have lasted minutes and, per the previous lesson, may have occurred on one venue. Third, and most important for comparison across assets, drawdown statistics on surviving assets are subject to survivorship bias: the assets that failed outright do not appear in the sample, so the observed distribution of recoveries is drawn from the population that recovered.
Volatility depends entirely on the window and the annualization
Realized volatility is the standard deviation of returns over a lookback period, usually scaled to an annual figure. Two window choices dominate the result. The lookback length determines how long a single extreme event stays in the number, so 30-day volatility and 90-day volatility can tell opposite-seeming stories for weeks after a dislocation, purely because one has dropped the event and the other has not. The annualization factor is a subtler trap. Because digital-asset markets trade continuously, daily returns are scaled by the square root of 365 rather than the square root of about 252 trading days used for equities. Comparing a figure annualized one way against a figure annualized the other introduces an error of roughly twenty percent before any real difference is considered.
Standard deviation also treats upside and downside dispersion identically, which is rarely what a reader means by risk. Downside deviation restricts the calculation to returns below a threshold and is closer to the intuitive concept. And volatility clusters: high-volatility periods follow high-volatility periods, so a recent reading is a better guide to next week than to next year, and any single figure implies a stability the series does not have.
| Figure | The question it answers | The question it cannot answer |
|---|---|---|
| Maximum drawdown | How far this asset fell from a peak within the window | How likely a similar fall is, or how long recovery would take |
| Realized volatility | How dispersed recent daily returns have been | The size of a tail event, which sits far outside the normal assumption |
| Correlation to bitcoin | How linearly two return series moved together in the window | Whether the relationship holds in stress, when correlations converge |
| Beta to bitcoin | The average magnitude of response to a move in the reference asset | Any causal direction, or stability of that response over time |
| Return divided by volatility | How much of a past return came with how much dispersion | Anything about the future, and it is not a Sharpe ratio unless a risk-free rate is subtracted |
Correlation is linear, windowed and unstable
A correlation coefficient measures the strength of a linear relationship between two return series over a chosen window. Three limitations matter here more than in most markets. It captures only linear co-movement, so two assets that move together violently in one direction and independently otherwise can show a moderate coefficient that describes neither state. It is highly sensitive to the window, and rolling correlations in this market swing across a wide range within a single year. And it rises toward one in stressed conditions, because forced selling driven by liquidation and margin calls is indiscriminate about what any individual asset does, which is precisely when a low historical correlation would have been relied upon.
Beta adds magnitude to correlation's direction, and inherits every one of those problems along with an additional one: it is a regression coefficient, and calling the reference asset an explanatory variable is a modeling convention rather than a statement about causation.
Time-based fields are not interchangeable
Several fields look like they answer the same question and do not. Days in drawdown counts time below the most recent peak, which resets whenever a new peak is set, while days since the all-time high is anchored to the highest observation in the entire recorded history and never resets until that level is exceeded. For an asset trading near its historical peak the two figures nearly coincide; for one far below it they differ by years and describe different things. Trend-reference fields have their own caveat: price relative to its 200-day average is a descriptive statistic about where a series sits against its own recent history, it is defined only once enough history exists, and a short-lived series produces a figure that is arithmetically valid and substantively empty.
The assumptions all three share
Each figure implicitly assumes the return distribution is stable and roughly well-behaved. Digital-asset returns are neither. They exhibit fat tails, so events that a normal distribution would call impossible occur at observable frequency, and any risk figure derived from standard deviation understates them. The history is short, and for most assets it covers a single sequence of macroeconomic conditions, so a full cycle of behavior has not been observed. Data series typically begin at listing rather than at creation, and the earliest period is often illiquid enough that the recorded prices reflect almost no trading. Backtests over such series are exposed to both look-ahead bias, when information is used that was not available at the time, and survivorship bias, when the sample is drawn from assets that still exist.
The reasonable use of these numbers is comparative and descriptive: what happened, over what window, relative to what else, with the window stated. The risk pages present the full set per asset with the windows labeled, compare places two assets on identical definitions, and the metric catalog gives the exact formula and lookback behind each field.