On this page
At a glance
- Formula
- t = mean ÷ (s ÷ √n), s = sample standard deviation
- Rough 5% threshold
- |t| ≈ 1.96 on large samples, two-sided
- Link to SQN
- SQN with population σ = t × √[n ÷ (n − 1)]
- Weakened by
- Dependence, fat tails, multiple testing
Formula and examples
Compute the mean of the trade results, their sample standard deviation s, and divide the mean by s ÷ √n. Sixty-four trades with a mean of +20 dollars and a standard deviation of 160 give t = 20 ÷ (160 ÷ 8) = 1.0: the average is one standard error above zero, which chance produces easily. Four hundred trades with the same mean and dispersion give t = 20 ÷ (160 ÷ 20) = 2.5. The strategy has not improved; the evidence has.
With many observations the t distribution is close to the normal distribution, so |t| above about 1.96 corresponds to a two-sided p-value below 5% under the model. With few trades, the threshold is higher.
Relationship with the SQN and the Sharpe ratio
A per-trade Sharpe ratio is mean ÷ standard deviation; multiplying it by √n gives the t-statistic. Van Tharp's system quality number uses the same construction, √n × mean ÷ standard deviation. Computed with the population standard deviation, as in the Replay statistics of Senzoukria, the SQN equals the t-statistic multiplied by √[n ÷ (n − 1)], a factor that is 1.005 at 100 trades. Reading an SQN is therefore reading a t-statistic, with the same caveats.
Why the number overstates certainty
- Trades are not independent when they cluster by regime or by session, which makes the true standard error larger than s ÷ √n.
- Fat tails make the standard deviation itself unstable on small samples.
- A t of 2.5 for the best of fifty configurations is not a t of 2.5 for a hypothesis stated in advance; the selection must be accounted for, for instance with a deflated Sharpe ratio.
In Senzoukria
The Replay report and the Performance panel give a verdict line built on the sample size and the SQN: under 20 trades the ratios are shown as exact but not significant; with a computable SQN, below 0 means no edge in the sample, below 1.6 weak, below 2.5 average, and 2.5 or more 'solid on this sample'. An SQN of 5 or more on fewer than 100 trades is flagged as suspicious, because a few trades are then carrying the result. The Gauntlet adds a probabilistic Sharpe ratio, a deflated Sharpe and a permutation p-value, which relax the normality and selection assumptions a plain t-statistic makes.
Related
In the same section
- Take-profit
- Systematic strategy
- Tape reading
- Survivorship bias
- Tape speed
- Strike
- Technical indicator
- Strike pinning
This page in other languages
Frequently asked questions
- What t-statistic should a strategy have?
- A value comfortably above 2 on a hypothesis defined before the test is the usual minimum for taking a result seriously. After a search over many configurations, a much higher bar or an explicit correction for the number of trials is needed.
- Is a t-statistic the same as a p-value?
- No, but they are linked: the p-value is the probability, under the null hypothesis of no edge, of a t at least as extreme as the one observed. The t is the distance; the p-value is the tail probability of that distance.