An analyst publishes a twelve-month call. A trend rule overrides it in weeks — adding on strength, cutting on weakness, both measured against SPY. Those two things disagree, and the disagreement is expensive. Koraki runs six books over one set of picks and measures which one the research actually survives.
OPEN THE WATCH →Every pick enters all six books on the same day at the same price. Nothing else is held constant, and nothing else needs to be — identical inputs mean the difference between the books is the rules and nothing but the rules.
A name enters the watchlist. Ticker and date go in. The entry fills at the next open, and the SPY level is stamped alongside it.
The same pick opens a position in every book at once. Four apply trend rules. One just holds, full size, for twelve months. The sixth records what actually happened.
Books add, stop, and exit at different moments on the identical price path. One books a gain near the high while another rides it back to the stop.
Each book is measured against the one that just held. Same pick, same path — so what’s left is what the rules cost or earned.
A test nobody will abandon isn’t a test. These four were written into the README before the first line of code, which is the only time they can be written honestly.
Thirty days without opening this page. Not looking is the cheapest kill signal there is.
Fewer than fifteen completed positions by month twelve. Too few observations to conclude anything, and no amount of patience fixes it.
At forty positions, the clustered 90% interval on fixed minus hold sitting entirely within ±$150 per pick per $10k and containing zero. An edge smaller than that gets ignored in practice anyway, which is what makes it the right threshold. A real finding: the rules don’t move the needle.
Missed entries, wrong timestamps, unadjusted splits. Better to restart clean than to publish a number built on a corrupt log.
Not on the list: a losing paper account. A negative result is a successful experiment with an unwelcome answer, and it gets published either way.