Put in the Sharpe you're assuming and how many variants you tried. There's no button to press. The results update as you type, and the defaults already show an answer below.
The result is a rough estimate of the total years of data needed: how much data it takes for a real edge of that size to pass the test with the probability you pick. The estimate assumes the true Sharpe never changes and that you test once, at the end.
Before you read the number:
The data you already have counts toward the total. This isn't how many more years you have to wait.
Having that much data doesn't mean the question gets settled.
Re-testing as new data comes in isn't covered, and neither is showing that something has no edge.
1. Years of data needed
A guess about the real edge, not a number from your backtest. A Sharpe measured over a few years is too noisy to use here. If you can't decide, try 0.2, then 0.5, then 1.0. To check a backtest Sharpe against the line instead, use section 2.
A whole number. Count everything, including parameter sets you dropped and date ranges you decided to ignore. If you're testing on data you didn't search on, count only the candidates you brought to it (1 if there's just one). If you never kept count, try 1 and 100 and see how far apart the answers are.
The highest false-alarm rate you'll accept: how often a strategy with no edge would still pass. A lower number is a stricter test and needs more years. One-sided means you only test for an edge in the profitable direction.
How often the test catches an edge that's really there. At 80%, it misses a real edge one time in five. A higher number needs more years.
2. The Sharpe your backtest has to beat
Uses the number of variants from section 1.
3. The whole table
Columns are variants tried (k). Rows are the true Sharpe. Each cell is the total years of data needed, rounded to a whole year. Changing the significance level or power redraws the table. The Sharpe and k boxes don't affect it, because those are already its rows and columns.
z is the inverse of the standard normal CDF. Give it a probability and it returns the point on the bell curve with that much area to its left. Dividing α by k is the Bonferroni correction: every extra variant makes the test stricter. It caps the family-wise error rate: the chance that at least one of the variants you tried looks like it has an edge when none of them do. That holds even when the variants are correlated. When they're close copies of each other, though, it's stricter than it needs to be.
The standard error of the Sharpe is taken as 1 ÷ √years, which assumes an annualized Sharpe built from daily returns. If you measure it from monthly or yearly returns, you need more years, and the gap grows with the true Sharpe. With iid normal returns the standard error is larger by a factor of √(1 + S² ÷ (2m)), where m is the number of observations per year (Lo, 2002). Real returns have fatter tails than a normal distribution and some dependence from one day to the next; those effects don't all push the same way. z is computed with Acklam's approximation.