What happened
In September 2026 we tested a long-short book of US small and mid caps: it picks stocks on their return over the past 12 months (skipping the latest month), rebalances every 63 trading days, and adds a risk brake that cuts exposure according to the state of the market. The rule we fixed in advance: take the net-of-cost Sharpe in each of two periods, 2006–2016 and 2017–2026, keep the weaker one, and pass only above 0.50. It scored 0.5589 and passed.
That number belongs to one rebalance start date. Our own backtest code starts on the first trading day by default and rebalances every 63 days from there. Starting on the second day, the third, and so on, with the same 63-day cycle, gives 62 more versions that are just as reasonable. Nothing makes the first day special.
Running all 63 start dates
Same strategy, same data, same costs; only the start date changes:
- The median is 0.3876, below the 0.50 bar
- The lowest is 0.0069 and the highest 0.7316, a spread of 0.7247, wider than the bar itself
- Only 15 of the 63 start dates clear 0.50
- The 0.5589 originally reported ranks 8th of 63
So we withdrew the claim that this book clears the bar. The same check showed that the risk brake improved the result at every start date in every period, and that part stands. Within one result, the part that is luck and the part that is real have to be read separately.
Why it happens
Rebalancing every 63 days means only a dozen or so rebalances in a period. The start date decides which days those rebalances fall on, and with them which periods the risk brake engages: in one of the periods, the number of rebalance periods in which the brake was active ranged from 1 to 6 depending on the start date. A few rebalances that happen to dodge or catch a sharp fall are enough to move the Sharpe of the whole period.
Whoever reports the result usually did not choose the start date; the code's default did. Because nobody chose it, nobody knows which of the 63 tickets they are holding.
How to check
- With an N-day rebalance cycle, run all N start dates and look at the distribution and the median, not just the first one
- Judge on the median, or at least state where the reported start date ranks among the N
- Or remove the start date altogether: open a new slice of the book every day, hold each slice for N days, and average the slices held at the same time. Done this way, the same book scores 0.4946, also below the bar
In our audits this is a key check for any strategy that rebalances less often than daily. It needs daily target weights and asset returns; when the client cannot supply them and the check cannot run, the report says so and the verdict is at most "conditional". It is never counted as a pass.