Trade Count and the Illusion of Independent Evidence

Imagine a hypothetical strategy that opens positions in twenty technology stocks on the same morning and holds them for five sessions. A broad sector move can affect all twenty outcomes. If another cohort opens the next morning, much of its holding period overlaps the first. Forty trade records are useful for execution review, but treating them as forty independent repetitions ignores their shared exposure.

That example does not imply a universal formula for dividing the trade count. Dependence can vary with the statistic, portfolio construction, holding period, and market environment. Begin by showing common entry dates, simultaneous positions, repeated instruments, and concentration in sectors or event clusters. A reviewer should be able to see where apparently separate observations came from.

A claim about implemented portfolio performance calls for a portfolio return series with contemporaneous positions, cash, and costs. A claim about an event response may require event-level outcomes and explicit handling of overlapping windows. A trade-level win rate answers a different question again. Moving between these units without explanation changes what is being estimated.

Aggregation is useful, but it is not a certificate of independence. Daily portfolio returns may remain serially dependent. Andrew Lo's work on Sharpe-ratio statistics shows why uncertainty and annualization depend on assumptions about the return process; the familiar square-root-of-time conversion cannot simply be applied to every correlated series.

An independent draw of individual trades can dismantle the time clusters that the study is meant to evaluate. A block bootstrap instead resamples stretches of a series to preserve some local dependence. The R boot package documents both fixed-length blocks and blocks with randomly varying lengths. Neither choice automatically captures every dependency or makes a changing market stationary.

For a strategy driven by common shocks, a practical review might compare uncertainty using chronological portfolio-return blocks and separately inspect outcomes by event cluster. Document the resampling unit, block-length rationale, random seed, and sensitivity to reasonable alternatives. These are study-design choices, not defaults to select according to whichever produces the narrowest interval.

More resampling iterations can reduce numerical variability in the simulation, but they do not supply additional historical market episodes. If most of the apparent edge comes from a small set of overlapping positions during one period, that concentration should remain visible even when the backtest contains many rows.

We would want a delivery to pair the headline estimate with its uncertainty method, exposure concentration, calendar coverage, and relevant limitations. Where independence cannot be defended, avoid translating a large trade count into a confident verdict. The useful next step may be a longer untouched evaluation or a narrower research claim, rather than another decimal place in the performance table.