Regarding city novelty

Hi metalhead,

it's the global novelty that matters. the way we will rank for counting quality submissions (and for allocating prize money) is given by a greedy algorithm that first ranks all submssions (whose sharpe ratios are statistically signficantly postive) by sharpe ratio , then finding the first one in the list that is no more than 50% time averaged cross sectional correlation with any of existing signals, then we add that to the list of existing signals, and recursively/inductively repeat until we can't find anymore of them. city novelty is a very compressed version of global novelty, because a city is just the normalized time average of the signal after it has been transformed using this thing called gauge transformation. so information is definitely lost, but, with good probability, if you are far from other cities (correlation<50%/degrees>60), you'll be far from other signals (that isn't 100% probability by the way). we can't expose the signals themselves obviously, so we give you city data in the release package. feel free to use that data in some way. you can also run python runner.py yourpredictor.py --gauge-fix locally to see how far you are (from a city perspective) to the signals we had prior to this contest before submitting your signal.

by the way: I note that the leaderboard is updated daily, so all metrics can and will change overtime, if a signal is borderline below or above "quality" it can flip..., which can change the count for quality submissions and hence the prize pot.

Thanks a lot. For my first submission, my city novelty came more than 60 deg and global novelty is at 84.9 Is it a good sign to refine that signal better? Or shift to a new approach. The way I am treating my 10 submission quota is like a portfolio diversification exercise with not similar signals but different signals to increase their odds of getting selected in the quality set. Is this the right mindset?

Yeah that’s the right mindset for sure . Your 80 degree global novelty submission is very good from a novelty a perspective . I’d suggest finding your next submissions to be high sharpe far from cities and far in signal correlation with your previous submissions. That way you have a chance of adding 10 quality submissions

I have a feeling a few users are kinda doing that already. There might be clever ways of using the city ambient coordinates data . Never tried myself but if I did I’d probably just get Claude Code to work on it lol

Yes, I have created and implemented multiple supplementary tools for the challenge. To fully and deeply understand all the data.

"Exactly. That's why I focused on building a robust local validation pipeline first. By building these supplementary tools, I can stress-test how features behave across different walk-forward windows without introducing leakage. Winning the leaderboard is secondary to building a structurally sound, uncorrelated signal that passes the proprietary novelty checks."

"Agreed. Having those custom tools is the only way to systematically find alpha in this setup. Now that the pipeline is solid, the next logical step is using that tooling to reverse-engineer feature spaces that naturally project far away from the crowded 'cities' like Distinguishingtremor, while keeping the CPU training under that 4-minute hard limit."