Sample size
What counts as one observation?
Bar count is not completed-trade count, while fills do not necessarily represent independent decisions. Define the analytical unit, measurement interval and position-grouping rules. Several pairs based on the same dollar signal can produce strongly dependent results. Duplicating observations or splitting an exit into separate records does not automatically add information about the future mean.
Count and precision
Under appropriate assumptions, mean SE = s/√n; multiplying n by four halves SE rather than dividing it by four. With estimated s, the normal-model example uses t with n−1 degrees of freedom, rather than automatically 1.96. See NIST: confidence limits for the mean. The interval concerns a model parameter; it does not contain 95% of future trades.
Why 100 is not a magic number
High variation, rare extreme losses and small effects require more information. Dependence and regime changes can invalidate a simple iid calculation. A larger sample does not repair misvaluation, leakage or choosing the best among many systems. The second example’s positive interval does not establish an edge if the strategy or period was selected after inspecting results. Account for the complete selection procedure.
Duplicated observations
Imagine 100 records that are exact copies of one random decision’s outcome. These are not 100 independent trials; duplication does not reduce uncertainty about the original draw in that model. Actual dependence is usually partial and must be estimated, rather than handled by arbitrarily dividing trade count by a chosen factor. Identify shared signals, overlapping positions and clustered events before selecting a sample-analysis method.
Use it in research planning
Specify the effect of interest, precision, horizon and conditions before collecting observations. Report sample count, dispersion and dependence; when resampling, check whether the method preserves relevant temporal structure. Do not stop a test arbitrarily at its first favourable moment or ignore missing data. In the course: strategy audit. Source checked on 11 October 2026.
