Three exits, four trading days on average
Consider three synthetic trades in different instruments. Each opens at the close of trading session 0. They close at the same time of day in sessions 1, 3 and 8. Count elapsed intervals between session closes, using the stipulated trading calendar.
| Completed trade | Entry session | Exit session | Elapsed trading days |
|---|---|---|---|
| A | 0 | 1 | 1 |
| B | 0 | 3 | 3 |
| C | 0 | 8 | 8 |
For completed trades with durations measured on the same clock,
An inclusive count of the entry and exit session labels would instead give 2, 4 and 9, averaging 5. That counts observed session dates, rather than elapsed session intervals. Either reporting convention needs a label; they cannot be compared as though they measure the same duration.
Nor are trading days automatically calendar days. In a hypothetical week with no intervening market holiday, Friday close to Monday close spans one trading-session interval and three elapsed calendar days. Intraday strategies need timestamps, time zones and session rules: rounding every same-date round trip to zero days can erase the duration being measured.
Decide what one trade means
A simple rule is to treat each instrument's continuous same-direction position as one trade: entry is the first fill that moves the position from flat, and exit is the fill that returns it to flat. A reversal closes the old trade and starts a new one. Under this rule, partial reductions do not end the trade or restart its clock. Alternatively, a report can match individual entry lots to exit lots and average their durations. That is a different observation unit and can produce a different answer from the same executions.
For a pairs trade, also say whether the observation is the whole paired position or each leg. Use actual fills if measuring realised exposure time; signal timestamps measure a decision horizon and may precede execution.
A closed-trade mean leaves an unfinished tail
Report the closed count and mean alongside the number and ages of positions still open. If selecting trades by exit date, use their actual entry dates even when those precede the reporting window; clipping entry to the window start shortens the measured holding period. An entry-date cohort answers a different question and may require later follow-up. Forced liquidation at the cutoff supplies an exit only if it is an explicit part of the simulated strategy. With no completed trades, the closed-trade average is undefined, rather than zero.
Duration is not dollars, frequency or turnover
The unweighted mean gives the smallest trade the same influence as the largest. For the three completed trades, stipulate entry notionals of USD 10,000, USD 10,000 and USD 40,000. The entry-notional-weighted mean is
That is a separate statistic, with entry notional as its chosen weight. It is not automatically a measure of committed equity, changing exposure or capital usage through time. Those questions need the relevant balance history.
Sources
- Marcos López de Prado, Advances in Financial Machine Learning (Wiley, 2018), Chapter 14, §14.3, “General Characteristics”: average holding period, bet grouping and its distinction from bet frequency and turnover. Publisher's book page.
- NIST/SEMATECH, e-Handbook of Statistical Methods, §8.1.3.1, “Censoring”: the distinction between an observed event time and an unfinished observation at a cutoff. The trade timeline and all numbers here are original synthetic examples.