Correlation Check

I am not able to understand how we are calculating correlation across competitions and within competitions, ideally If I have a better sharpe alpha ideally despite of correlation they should be considered for scoring.
Example :
Suppose I submitted a very bad sharpe , unpolished submission with very low sharpe, but in subsequent competitions I improved it and sharpe was raised significantly , then it should be considered, else its a loss for both the platform and individual . And this is a common practice across various quant competitions like if you improve sharpe by say 10% compared to previous submission, it should be considered.
My signal_7b2787a70db19dd8 has the highest sharpe across both competitions , more than 10% compared to any previous submission by anybody, still ist not considered.

On the other hand I see a lot of signals by user " braxtonmensah" which has almost same city novelty, still considered in the qualified signals .
This is very hard to understand and hence not able to iterate further , seems like a coin toss or luck whether your signal will be qualified or not.

Hi departinginkwell,

Thanks for laying it out — and first, congratulations: signal_534b9cce8e01e809 is Q2 in Cycle 2 right now. I'll come back to why that one qualified and the other didn't, because it's the whole answer.

There are three things that are easy to run together here.

Sharpe sets the order. Correlation decides admission. City is a map, not the test. Within a cycle, statistically significant positive-Sharpe signals are sorted highest first. Each is admitted only if its time-averaged cross-sectional correlation is at or below 0.5 in absolute value against every signal already occupying space — earlier admits in the same cycle, the season's legacy pot, and our background signals. City distance is a navigation aid for that test: far apart almost always means uncorrelated; close may or may not be. The correlation itself is what decides.

On improving an alpha. We genuinely appreciate it — pushing a signal further in the same direction is real work, and it's the work that makes a signal worth owning. But it isn't what the cycles are set up to reward. When we move from one cycle to the next, accepted signals carry forward into the season's legacy pot and keep their place; we tried to be clear about that in the rules, and the note that a resubmission blocks itself at correlation ~1 is the same point from the other side. So the question each cycle asks isn't "can you improve the alpha you found?" — it's "having found that one, can you find another?" A library of fifteen polished versions of one predictor is worth less to a portfolio than fifteen different predictors, even if every polish adds 10%. That's a deliberate trade-off: exploration during the season, and the pot clears at the season boundary.

What happened to signal_7b2787a70db19dd8. Two corrections to the premise. It isn't the highest-Sharpe signal in the pool — it's second, and the top one is Q1 above you. And it's very close to your own signal_8329109176f23205: 2.9° of city distance and a measured correlation of 0.62 between them — the same approach, twice. What excluded both is a signal accepted in Cycle 1, signal_7acec9c282adae08: yours correlate with it at 0.53, just over the line. That's the carry-forward rule doing exactly what it says.

Your 534b… is the counter-example: a different approach, correlation 0.42 to that same Cycle-1 signal, and it's in the set at #2.

On braxtonmensah. The two qualified signals sit at 53° and 18° from their nearest cities — not comparable to 2.9°. The others aren't qualified; they have negative Sharpe. Similar-looking novelty numbers don't imply similar correlation relationships, and close cities are exactly where the test bites.

And the part that matters most for you. A signal that doesn't make the quality set in a given cycle isn't lost. The improved 7b27… is a strong signal and we can see that. Our revenue-sharing scheme onboards signals on the basis of PnL attribution — what a signal actually contributes to the book — not on cycle placement. I'm not going to make promises about timing here; onboarding has its own process. But "I improved my alpha" is precisely the thing that route rewards, where the cycle prizes reward "I found a new one."

It isn't a coin toss. --gauge-fix is the tool: build, check distance to the occupied cities, move, repeat. Your 534b… is what that looks like when it works, and it's the one we'd like more of.

If it would help to walk through any of this live, I'm happy to hop on a call — just say the word.

Best,
NONIUS

Correlation Check

3 Replies