Скачать с ютуб Numerai Quant Club with Michael Oliver в хорошем качестве

Numerai Quant Club with Michael Oliver

Join Numerai’s Chief Scientist and Minister of Research, Michael Oliver, as he digs deep into the world of Numerai, research, data and everything in between. This will be a recurring event through 2023. More details: https://forum.numer.ai/t/numerai-quan... [02:07] What datasets and feature types for Signals would complement Classic? Eg price/volume short term reversal features? Does Yahoo data have future leakage? [03:22] What are your thoughts on lagged targets as features in Signals? [03:51] Could we get a target that if submitted would have given optimal TC (given the meta model in that era) Might be even more useful than a meta model prediction file [04:50] What's the latest thinking about TC as a feedback mechanism? If TC doesn't predict future TC (much), what good is it? [06:58] Not sure the right place. Could we get corr with the meta model in diagnostics for the regular tournament? Low corr with the metamodel seems to help with TC. [07:25] How can we optimise TC Sharpe? [09:14] How to determine TC leverage? Can we get different TC leverage levels simulated on diagnostics? [09:35] Can you run us through how a hypothetical model with lots of exposure to feature X (relative to target Y) goes through the pipeline until TC is calculated? [12:56] As I understand it, the different targets are neutralized against things like sector, or currency - which targets are the "least" and "most" neutralized? [15:08] What are your thoughts on CNN on IGTD with respect to getting TC? [23:15] Could you tell us which targets are least exposed to momentum? [24:00] Does Nomi best reflect what the optimizer is doing with our ranked_pct/normalized/stake_weighted/ranked/normalized/exp(1.5)/stdev_1'd predictions? [26:47] Will you be releasing any more ‘real’ new features in the near future? Other than transformed or synthetic features? [28:11] Often a single day change in TC (same round) goes from 90ish percentile to less than 20 pct or reverse. How is it so sensitive to big RANK changes day-to-day? [37:05] Is corrWMM the best indicator of uniqueness given many agree on the middle portion [38:09] How do small changes in the optimizer impact TC? [39:49] Let's say you do something to reduce TC volatility. Will my current scores be any guide to my future scores under altered/smoothed scheme? [45:40] Why does the meta model parquet file only contain 150 eras? Also, I was under the impression that it'd be updated every week, and AFAIK, it wasn't updated last week. [45:53] How can we locally compute metrics post applying the optimiser. Sounds like 2 steps of black boxes rather than 1 currently [48:09] Will we get daily targets? [49:15] When staking is enabled for daily, are we still gonna be able to do that thing we talked about where late submissions can be rolled-over/staked for the next day? [50:30] Targets Thomas, William, and Arthur have different bin numbers than the other targets. Eg 7 instead of 5 bins. Are they the same as eg Nomi, Jerome, etc, just more granular? If so which ones? [54:29] NumerCon Want to join in on the hardest #datascience competition on the planet? Or are you a quant with your own data? https://rocketchat.numer.ai https://docs.numer.ai/tournament/learn Numerai: https://numer.ai Numerai Signals: https://signals.numer.ai