Quant Analytics: from scratch to pro · Statistics for markets · Intermediate
How sure is an average? Standard error and confidence intervals
A mean return measured from a sample is an estimate with a margin of error. The standard error shrinks only with the square root of the number of observations, which is why telling a small edge from zero can take decades of data.
In this lesson
- Compute the standard error of a mean and a 95% confidence interval
- Work out how many years of data are needed to separate a given edge from zero
- Compute the standard error of a proportion such as a hit rate
- Explain why sampling more often does not add information about the mean
Read first: Averages that mislead: arithmetic and geometric means, Volatility: from daily moves to an annual figure
A backtest or a track record gives a number: a mean return of 6% a year. The honest follow-up is how far that number could be from the truth. A sample mean is an estimate, and the margin around it is usually wide for financial data.
The standard error
If n observations each have standard deviation sigma, and they are independent, the mean has a standard error of
SE = sigma / sqrt(n)
The standard error is the typical size of the gap between the sample mean and the true mean. It falls with the square root of n: four times the data halves it.
A confidence interval turns it into a range. A 95% interval is mean +/- 1.96 x SE. If the same procedure were repeated on many fresh samples, about 95% of the intervals would contain the true mean. It does not say the true mean is inside this particular interval with 95% probability, though that is how it is often read.
The t-statistic is the mean divided by its standard error: t = mean / SE. It counts how many standard errors the mean is away from zero.
Worked example
Ten years of annual returns have a mean of 6% and a standard deviation of 15%.
- Standard error:
0.15 / sqrt(10)= 4.74%. - 95% half-width:
1.96 x 4.74%= 9.30%. - Interval: 6% minus 9.30% to 6% plus 9.30%, which is -3.3% to 15.3%.
- t-statistic:
6% / 4.74%= 1.26.
Zero is inside the interval. (With only ten observations the t-distribution multiplier of 2.26 replaces 1.96, which widens it to about -4.7% to 16.7%.) A decade of data cannot separate a 6% edge from nothing, even though 6% a year looks attractive on a chart.
How many years are enough
Set t = 1.96 and solve for n: n = (1.96 x sigma / mean)^2. With a mean of 6% and a standard deviation of 15%, that is 24.0 years.
In terms of the ratio of mean to volatility, n = (1.96 / ratio)^2. A ratio of 1.0 needs 3.8 years. A ratio of 0.2 needs 96 years. The chart shows the curve. Most real edges sit at the low end of the ratio, which is why long samples, many independent markets or a prior reason for the effect are needed. See the Sharpe ratio, whose statistics are treated in Lo (2002).
More data points is not more information
Cut the same ten years into 120 months. The monthly standard deviation is 15% / sqrt(12) = 4.33%. The standard error of the monthly mean is 4.33% / sqrt(120) = 0.395%, and scaled to a year (times 12) it is 4.74% again.
Sampling more often does not improve the estimate of the mean. Only a longer span, or more independent series, does. It does improve estimates of volatility, which is why volatility is measured so much more precisely than the mean.
Proportions: hit rates
A hit rate is a share of trades that win. Its standard error is sqrt(p x (1 - p) / n). A rule that wins 55 of 100 trades has an SE of 4.97%, and its 95% interval is 45.2% to 64.8%. Fifty percent is inside, so luck is a live explanation.
To separate 55% from 50% at this level of confidence takes (1.96 x 0.5 / 0.05)^2 = 384 trades.
When the formula is too kind
- Serial correlation. The formula assumes independence. Positively autocorrelated returns carry less information than their count suggests, and the true SE is larger.
- Heavy tails. A few extreme days dominate the sample standard deviation and make it unstable. The bootstrap lesson shows an alternative.
- Small samples. With few observations use the t-distribution, whose multiplier is larger than 1.96 (about 2.26 for ten observations).
- Selection. If this was the best of many tries, the interval is wrong before you start. See multiple testing.
Common mistakes
- Reading a mean without its standard error.
- Reading a 95% interval as a 95% probability for the true value.
- Adding data points by sampling more often and believing the mean is better known.
- Using 1.96 for a sample of ten.
- Treating 55 wins out of 100 trades as proof of an edge.
In practice
Report every average with its standard error or interval, and state how many independent observations stand behind it. Ask how many years or trades the claim would need, and compare that with what exists. Research on FinUties is where studies built on methods like these are published, with their method and limits stated.
Years of data needed for a t-statistic of 1.96
Worked numbers
Each result is recomputed from the formula shown.
| What | Formula | Result |
|---|---|---|
| annual standard deviation | 0.15 | |
| annual mean return | 0.06 | |
| standard error of the mean, 10 years | sigma / sqrt(10) | 4.74% |
| 95% half-width, 10 years | 1.96 * se10 | 9.30% |
| lower end of the interval | mu - half10 | 3.3% |
| upper end of the interval | mu + half10 | 15.3% |
| t-statistic of the mean, 10 years | mu / se10 | 1.26 |
| years for a t-statistic of 1.96 | (1.96 * sigma / mu)**2 | 24.0 |
| years needed at a ratio of mean to volatility of 1.0 | (1.96 / 1.0)**2 | 3.8 |
| years needed at a ratio of 0.2 | (1.96 / 0.2)**2 | 96 |
| monthly standard deviation | sigma / sqrt(12) | 4.33% |
| standard error of the monthly mean, 120 months | sd_month / sqrt(120) | 0.395% |
| same standard error scaled to a year | se_month * 12 | 4.74% |
| standard error of a 55% hit rate over 100 trades | sqrt(0.55 * 0.45 / 100) | 4.97% |
| lower end of the hit-rate interval | 0.55 - 1.96 * se_p | 45.2% |
| upper end of the hit-rate interval | 0.55 + 1.96 * se_p | 64.8% |
| trades needed to separate 55% from 50% | (1.96 * 0.5 / 0.05)**2 | 384 |
Check yourself
Ten years of annual returns have a mean of 6% and a standard deviation of 15%. What is the standard error of the mean?
0.15 / sqrt(10) = 4.74%.
What is the 95% confidence interval for that mean, and does it include zero?
6% +/- 1.96 x 4.74% gives about -3.3% to +15.3%. It includes zero, so ten years do not rule out a true mean of zero.
How many years of data does it take for a mean of 6% with a standard deviation of 15% to reach a t-statistic of 1.96?
n = (1.96 x 0.15 / 0.06)^2 = about 24 years.
A rule wins 55 of 100 trades. Is a 50% hit rate ruled out?
The standard error is sqrt(0.55 x 0.45 / 100) = 4.97%, so the 95% interval is about 45.2% to 64.8%. 50% is inside it, so it is not ruled out.
You switch from annual to monthly returns. Does the standard error of the annual mean improve?
No. The 120 monthly observations have a standard deviation of 15% / sqrt(12) and give the same 4.74% for the annual mean. The span of time sets the information, not the number of data points.
Key terms
Standard errorConfidence intervalT-statisticStandard deviation
Sources
- The Econometrics of Financial Markets. Princeton University Press (John Y. Campbell, Andrew W. Lo, A. Craig MacKinlay), 1997. Accessed 9 Nov 2026
- The Statistics of Sharpe Ratios. Financial Analysts Journal 58(4), 36-52 (Andrew W. Lo), 2002. Accessed 9 Nov 2026