Drawdowns and Recovery: Depth Is Only Half the Story

Choose a legitimate value series

Drawdown analysis begins with an account-value or wealth-index series whose construction is understood. A futures reference price is not automatically an investment wealth index: it can change contract identity and does not include account funding or implementation. Similarly, a sequence of directional states has no investment drawdown on its own. Before using the term, identify precisely what the plotted values represent and what they exclude.

For the following arithmetic, use an invented positive wealth index without external deposits, withdrawals or charges. Its values are teaching inputs, not a record of a strategy or account. This narrow setup makes the definition visible. Real reporting requires an explicit treatment of cash flows, fees, valuation times and missing observations before the same formula can be interpreted as an investment experience.

Calculate against the running peak

Let the hypothetical index be 100, 110, 99, 88, 96.8 and 110 at observations zero through five. Its running peak is 100 initially and 110 thereafter. Express drawdown as a positive magnitude: one minus current value divided by the running peak. At 99 the drawdown is 10%; at 88 it is 20%; at 96.8 it is 12%. The maximum observed drawdown is 20%.

A running maximum is a standard cumulative operation documented in the pandas library. The documentation also makes missing-value handling explicit, which matters when implementing this definition. A computational running peak does not by itself solve economic measurement problems. The index still needs chronological ordering, consistent units and valid valuations. A correct cumulative calculation on an inappropriate source series can produce an inappropriate interpretation.

Recovery arithmetic is asymmetric

From 88, regaining 110 requires an increase of 22. Relative to the depressed value, that is 22 divided by 88, or 25%. A 20% decline therefore does not require merely a 20% increase to recover. In general, for a decline fraction d below one, the required increase is d divided by one minus d. A 50% decline requires a 100% increase.

The denominator explains the asymmetry: the decline is measured from the earlier peak, while recovery is measured from the lower starting point. This arithmetic supplies no forecast of the time or likelihood of recovery. A table of required increases should not be presented as a set of achievable targets. It describes the mathematical distance to a previous value, not a plan for getting there.

Define the duration convention

In the invented path, the peak occurs at observation one and is regained at observation five. Peak-to-recovery duration is four observation intervals. There are three strictly underwater observations: two, three and four. Trough-to-recovery duration is two intervals. All three counts are correct for their definitions, but they are not interchangeable. State which clock is being reported before comparing durations across studies or accounts.

Calendar time and observation time can differ because of weekends, closures and missing records. An episode ending at the last available observation below its peak is unrecovered, not a completed episode with a known recovery date. Such episodes should remain visible in summaries. Dropping them because their full duration is unknown makes the historical experience look shorter and more complete than the evidence supports.

Sampling frequency can conceal depth

Suppose a more frequent valuation falls to 80 between the reported observations of 88 and 96.8. Against a peak of 110, that additional point gives a drawdown of 27.2727%, larger than the 20% seen in the coarser sequence. This hypothetical difference does not imply that the coarse report is arithmetically wrong. It shows that the report measures only what its sampling frequency captures.

Use consistent valuation frequency when making comparisons, and identify whether the values are intraday, session-end or periodic statements. A monthly series cannot establish the maximum intramonth decline. Conversely, adding noisy or incomparable intraday marks can create its own problems. The appropriate interpretation follows the valuation definition and evidence quality, not simply the finest timestamp that happens to be available in a dataset.

Separate a drawdown from a diagnosis

Rej, Seager and Bouchaud study drawdown depth and length under an upward-drifting Brownian-motion framework. Their paper illustrates why an adverse run alone does not automatically distinguish a deteriorating strategy from variation within an assumed process. The framework's assumptions matter. It does not supply a universal threshold beyond which any real strategy is broken, nor does it justify dismissing every adverse episode as ordinary noise.

A diagnostic review should separately examine data integrity, unintended exposure changes, costs and deviations from the declared method. These are operational explanations that a drawdown statistic cannot resolve. Avoid choosing a new stop threshold after inspecting the exact episode and then describing it as a pre-existing control. Any comparison of alternative responses needs its own declared question and cannot retroactively change the original record.

Do not let cash flows manufacture recovery

An external deposit can raise account equity above an earlier level without reversing investment losses. In a simplified example, adding $20,000 to an account that fell from $100,000 to $80,000 restores its cash balance to $100,000, but the deposit is not investment recovery. A properly constructed performance index handles external flows explicitly instead of counting every increase in account value as a return.

A useful drawdown summary therefore includes the value-series definition, flow treatment, frequency, maximum depth, duration convention and current recovery status. Preserve the underlying episode boundaries so another reader can reconstruct the arithmetic. Depth communicates how far a path fell; duration communicates how long it remained below its peak. Neither alone captures liquidity pressure or proves how an investment will behave after the sample ends.

Source record

1. Rej, Seager and Bouchaud: You Are in a Drawdown. When Should You Start Worrying?
2. pandas: Series.cummax and missing-value behaviour