A crypto backtest has no universal minimum trade count. The useful sample size depends on the metric, outcome variability, dependence between observations and the historical conditions covered. Count completed positions consistently, preserve shared-event and overlapping-exposure groups, and estimate uncertainty with a method that fits those dependencies. Hundreds of trades from a few market episodes can provide much less independent evidence than the headline count suggests.
The question to ask is more specific than “Have I reached 100 trades?” It is: “What am I estimating, which observations supplied it, and how much new information did each observation add?”
Define one observation before counting it
Ten partial fills of one entry are not ten independently tested trading decisions. Nor do five partial exits necessarily constitute five complete strategy campaigns. Keep fill identifiers, position identifiers and any parent campaign identifier so another reader can reconstruct the reporting unit.
For a mean net outcome per completed position, aggregate the entry, additions, exits and attributable costs into that position. Retain unfinished positions separately; excluding them from a completed-trade statistic does not remove their exposure from portfolio equity. The trade expectancy and win-rate guide covers that accounting. This article asks how much independent evidence the resulting observations contain.
Choose the target metric first. Mean dollars per position, mean outcome divided by initial planned risk, win fraction and mean daily portfolio return are different estimands. They can have different dependencies and effective sample sizes. A precise estimate of win frequency does not automatically give a precise estimate of a payoff distribution with rare large losses.
Keep 120 trades fixed and change their dependence
Consider an original hypothetical model: 12 independent market episodes each generate 10 completed positions, giving 120 observations. Position outcomes share a common mean and the same marginal variance. Within an episode, every pair of outcomes has the same correlation, rho; outcomes from different episodes are independent. All observations receive equal weight in the mean.
Those are strong assumptions chosen to expose one mechanism, not observations about crypto prices or KLINEO results. Here the variance of the mean, compared with 120 independent observations of the same marginal variance, is multiplied by:
D = 1 + (m − 1) × rho, where m = 10.
The variance-equivalent independent sample size is N_eff = 120 / D. The standard error is multiplied by sqrt(D).
| Assumed within-episode outcome correlation | Variance multiplier D | Effective observations for this mean | Standard-error multiplier |
|---|---|---|---|
| 0.00 | 1.00 | 120.00 | 1.00 |
| 0.20 | 2.80 | 42.86 | 1.67 |
| 0.60 | 6.40 | 18.75 | 2.53 |
| 1.00 | 10.00 | 12.00 | 3.16 |
The 0.60 row follows from 1 + 9 × 0.60 = 6.40, then 120 / 6.40 = 18.75. J-PAL describes this design-effect expression in its guidance on clustered evaluation designs. Its subject is randomized evaluations, not trading validation; the table uses the stated covariance model as an educational analogy.
At correlation 1, each group contributes ten identical outcome observations, so the 120 rows contain only 12 independent outcome groups under this model. At 0.60, “18.75” describes equivalent precision for the mean. It does not delete 101.25 trades, set confidence-interval degrees of freedom, or certify the quality of any strategy.
Do not insert a price correlation between BTC and another coin into this table. The needed quantity is the correlation of the measured strategy outcomes under the chosen grouping model. Unequal group sizes, unequal variances, changing position weights and dependence across episodes require different calculations.
Look for shared causes in the ledger
Group identifiers should describe a rule a reader can reproduce. Examples include positions opened by one signal timestamp, overlapping holdings from one campaign, or a predefined event window. Assign those rules before inspecting which grouping makes the result look strongest.
A failure case is a scanner that opens ten related positions during each of twelve sharp market moves. Its report shows 120 completed trades, but a shared shock may drive many outcomes together. Adding another symbol during those same episodes can add exposure without adding as much independent evidence as a later, distinct episode. The table illustrates that possibility; the actual dependence must be investigated.
Inspect entry and exit times, overlap, common signals and concentration by episode. For a portfolio question, a consistently sampled marked-return series may be more suitable than sorting irregular position outcomes by exit time. Reordering simultaneous exits by symbol can manufacture a misleading adjacent-row pattern.
NIST's autocorrelation definition assumes equally spaced observations. An irregular trade list needs an explicit interpretation of its lag; “one trade later” is not the same elapsed time throughout the sample. NIST also explains that an absence of significant autocorrelation does not establish randomness in its autocorrelation-plot guidance. A quiet lag-one plot cannot prove independence across common market shocks.
State the precision question before requesting more trades
For independent outcomes with a common finite variance, the standard error of their mean scales as 1 / sqrt(N). NIST's autocorrelation-plot guidance warns that the familiar s / sqrt(N) calculation relies on the randomness assumption. With the model otherwise unchanged, four times as many independent observations halves that standard error. Four times as many duplicated or strongly related rows need not.
Write down the metric, tolerable estimation uncertainty and dependency model before setting an observation target. A highly variable net payoff can require more evidence to estimate its mean precisely than a tightly concentrated payoff. An observed sample with no large losses does not prove the loss tail is absent.
Where a method resamples observations, its resampling unit must match the proposed dependence structure. Resampling individual trades breaks the shared-episode relationship in the hypothetical example. Resampling groups preserves that particular relationship, but only twelve groups still leave limited information, and their assumed independence may fail. Do not present a computed interval as evidence that those assumptions are true.
More observations cannot repair missing evidence
Sample size concerns uncertainty within an observation process. It does not resolve incorrect fills, missing assets, selection bias or future-informed inputs. A million rows generated using tomorrow's information are still compromised observations.
More granular candles also have a different purpose: they can resolve some sequence ambiguity within a trade. They do not automatically create more completed independent trade outcomes. The intrabar backtest-data guide separates finer historical sequence evidence from the execution details candles still omit.
Historical coverage is another dimension. A long series confined to similar conditions may estimate that particular period more precisely without demonstrating behavior in an unobserved liquidity shock or different market structure. Describe the dates, instruments, regime definitions and concentrations actually tested. Calling twelve groups “independent” cannot make them representative of every future condition.
For a process that refits over time, walk-forward testing adds a chronological evaluation protocol. Repeatedly changing that protocol after inspecting its results changes what the observations can evaluate.
Trade counts also say little about account-level losses. Use marked equity, keep open exposure visible, and apply a consistent cash-flow treatment. The cash-flow-adjusted drawdown guide explains why an added deposit must not erase a prior portfolio decline from the measurement.
A sample-size review card
Attach this card to the frozen run rather than publishing only a total trade count.
| Review field | Evidence to retain | What remains unresolved without it |
|---|---|---|
| Target quantity | Metric, units, weights and treatment of unfinished positions | What the observations are intended to estimate |
| Counting unit | Fill-to-position and position-to-campaign mapping | Whether bookkeeping inflated the denominator |
| Dependence | Timestamps, overlap groups and declared episode rule | Whether ordinary independent-observation uncertainty fits |
| Concentration | Counts and exposure by episode, instrument and period | Whether one event dominates the report |
| Precision method | Assumptions, grouping or sampling interval, and sensitivity choices | What a reported effective size or interval actually means |
| Evaluation boundary | Frozen settings and complete trial history | Whether selection influenced the supposedly evaluative sample |
Keep the backtest trial log alongside this card. Dependence and strategy selection are separate issues: adjusting for one does not correct the other. A reproducible conclusion might be “120 completed positions across 12 predefined episodes; independent-trade precision is unsupported.” That statement gives the next research task a clear scope.
Method and limits
This article is AI-assisted educational synthesis of the linked primary statistical guidance. The equal-group model, table and review card are original illustrations; the arithmetic was recomputed. No historical trade ledger, dependence estimate, confidence interval or live execution experiment was produced. The source examples establish statistical concepts within their own settings. Applying them to a trading record requires a suitable model and does not establish future profitability.