One hundred return observations do not automatically provide one hundred independent pieces of evidence. This lesson derives how dependence changes uncertainty in an average and shows how to question a historical performance claim without mistaking a precise calculation for a reliable forecast.
A return is the gain or loss over a stated interval, expressed relative to the starting value. The arithmetic mean adds the returns and divides by their count. Volatility, measured by standard deviation, describes the spread of individual returns. Variance is standard deviation squared. Standard error describes something different: how much an estimated mean would vary across repeated samples from the same process.
Independent observations supply no information about one another. Serial dependence means observations are related across time; autocorrelation measures one linear aspect of that dependence. A sample is the observations you collected, while a process is the assumed mechanism generating them. Sample correlations estimate process correlations with error.
Suppose a report contains 100 daily returns. First ask whether these are consecutive one-day returns or overlapping multi-day returns reported every day. Two adjacent five-day returns share four days. Their rows reuse underlying observations, so treating the rows as independent can exaggerate the evidence even if the underlying daily returns were independent.
The glossary calculates lag-one autocorrelation from four illustrative returns. The same numbers in another order have the same mean and volatility but a different lag-one statistic. Counting rows or inspecting a histogram, which groups values without retaining their order, misses that distinction.
Assume a weakly stationary process: the mean and variance stay constant over time, and covariance depends only on the lag. Covariance measures whether two deviations from their means tend to have the same sign. Write σ² for the process variance and ρₖ for its autocorrelation at lag k. These are assumed process quantities here, not values established by a short sample.
For a sum, add each observation’s variance and twice the covariance of every distinct pair. At lag k there are n − k pairs. Dividing the sum by n to form the mean divides its variance by n². Therefore Var(R̄) = (σ² ÷ n) × [1 + 2 Σ(k=1 to n−1)(1 − k/n)ρₖ]. The standard error is the square root of this variance.
If observations are independent, every ρₖ is zero and the bracket is 1, giving standard error σ ÷ √n. Positive covariance terms enlarge the bracket and uncertainty in the mean. Negative terms can reduce it. What matters is the weighted contribution of all relevant lags, not simply the sign of lag one.
Define effective sample size for this mean as n divided by the bracket above. It is the number of independent observations with the same marginal variance that would give the same variance of the mean. It can be fractional. It is a comparison of precision under a model, not a count of unique dates or a score for the dataset.
Illustrative model: assume 100 daily returns, daily standard deviation σ = 1 percentage point, and process autocorrelations ρₖ = 0.5ᵏ at every positive lag. This geometrically decaying pattern is an assumption, not something implied merely by observing a lag-one estimate of 0.5.
For a long sample, approximate the bracket by 1 + 2(0.5 + 0.25 + 0.125 + …) = 3, because that infinite sum is 1. Then effective sample size is approximately 100 ÷ 3 = 33.3, and standard error is approximately 1 ÷ √33.3 = 0.173 percentage points. Under independence it would be 1 ÷ √100 = 0.100 percentage points. Keeping the finite-sample weights gives a bracket of about 2.96 and standard error of about 0.172 percentage points, so the approximation is close in this assumed example.
The daily volatility remains 1 percentage point in both models. Dependence changed uncertainty in the average, not the assumed spread of each day’s return. The Stan manual presents the related long-sample effective-size formula for correlated simulation draws; this lesson applies the variance-of-a-mean identity to an explicitly assumed return process, without treating financial returns as simulation output.
Real process correlations are unknown. A short sample gives noisy estimates, especially at long lags where fewer pairs remain. Changing economic conditions can invalidate stationarity. Heavy tails, meaning unusually frequent extreme outcomes, can also make estimated uncertainty unstable. Do not insert a short-sample lag-one estimate into the geometric model unless that model has separate support.
Analysts can use dependence-aware standard errors or a block bootstrap, which resamples consecutive groups instead of isolated observations. Their lag choices, block lengths and stability assumptions still need justification. Randomly shuffling return rows destroys their time order; it does not establish that the original observations were independent.
Lo and MacKinlay’s introduction explicitly separates rejection of a random-walk price model from a conclusion about market inefficiency. Likewise, an effective sample size neither establishes a positive true mean nor demonstrates a tradable pattern. Future evaluation, costs and changing conditions remain separate questions.
Illustrative assessment: suppose n = 120, σ = 2 percentage points per day and the same assumed ρₖ = 0.5ᵏ pattern. Using the long-sample approximation, what are the effective sample size and standard error of the average daily return? Has the daily volatility changed? Could a sample lag-one estimate alone justify this calculation?
Answer: the bracket is approximately 3, so effective sample size is 120 ÷ 3 = 40 and standard error is 2 ÷ √40 ≈ 0.316 percentage points. Under independence it would be 2 ÷ √120 ≈ 0.183 percentage points. Daily volatility remains the assumed 2 percentage points. Lag one alone cannot establish the remaining correlations or stationarity, so it cannot justify the calculation.
A second report says its lag-one correlation is near zero and therefore its returns are independent. Is that conclusion supported? Answer: no. Dependence at other lags or nonlinear dependence may remain. An absence of detected linear correlation is weaker than evidence of independence, particularly in a short sample.
General education, not investment advice. Every numerical example and assessment is illustrative, not a market estimate or a finding about Quantustik. The calculations assume a stable process with finite variance. Effective sample size does not measure strategy quality or guarantee future performance.
Yes, for some dependence patterns negative covariance reduces variance of the mean below the independent-observation benchmark. This is specific to the quantity being estimated and does not establish better financial performance.
No. Repeated or overlapping information is not equivalent to new independent evidence, and a changing process can make old observations less relevant. More data helps only within defensible assumptions.
No. It describes uncertainty in an estimated mean under a sampling model. Tomorrow’s individual return has its own uncertainty and can be far from that mean.
Investor education only — not investment advice, and never a promise of profit. Every investment can lose value.