True Contribution for dummies

[anthill]

@autratec

I don’t quite understand the logic and removing and add in that signal will help dramatically improve the portfolio performance, considering CORR/IC has such low score.

To give a heuristic example of how you could see a situation where a model gets very low Corr but very high TC, imagine that the market is divided into a number of different sectors and the metamodel does extremely well on all but one of these sectors. If you provide a model which makes very good predictions on that one missing sector, but random predictions on all the other sectors, you will end up with a very low Corr, but your model will nevertheless dramatically improve the performance of the metamodel and will get a high TC.

[wigglemuse]

Yes, some models can be considered like vitamins for the metamodel – they are just helping supplement some deficiency.

[autratec]

thanks for the explanation. it helps me thinking special strategy to get higher TC, rather than general CORR.

[f58c]

@wigglemuse - I’ve uploaded some models and they are now getting tc scores. from day to day, past rounds tc changes and updates. is tc in flux until rounds are “resolved”? my best models tend to stay relatively positive even with these changes, but I’m using tc scores to judge how much NMR to stake. any insight is greatly appreciated! thx!

[wigglemuse]

All scores on unresolved rounds are in flux – only the final day (day 20) of the round actually counts. The in-progress scores are “as-if” it ended today and are just something to look at.

[f58c]

thx! my earliest models will resolve at the end of Sept. it’ll be interesting to see the results.

[numerologist]

is tc in flux until rounds are “resolved”?

That’s something I don’t understand myself too.
One would expect that values would stabilize closer to the end of a round: the farther the round from resolution, the more volatile its TC value (similar behavior we have with CORR).

But I checked one of my models which is basically 1-p example predictions, and on Thursday night it had (expectedly) horrible cumulative CORR but a very attractive (literally all the way up) cumulative TC.
However, when I checked it on Friday morning, the result for CORR was about the same but all TC values for all the unresolved rounds literally flipped, i.e. it became cumulative negative TC all the way down.

How is something like this even possible overnight? It feels like something is really off (a bug?) about the algorithm that computes TC.

[gammarat]

@wigglemuse

Leaderboard position is score based only – it used to be corr, now it is TC. (20 round moving average with the current live rounds gaining in weight each day while the 4 least recent of 20 lose weight each day – at least that’s how it was done on corr).

The leader board is now based on a 1 year average, and it can be ranked on any of the headings. It just comes up first on TC.