This is an episode on the super cool Machine-Learning-powered hedge fund, Numerai.
Numerai is a global equity long-short hedge fund, which specializes in relatively long time-range investments, on the scale of 6-9 months. They’re interested in more finely granular moves, that are at the level of individual stocks; they aim to identify anomalously-priced stocks, and thus exploit price discrepancies. One might argue that their essential task is to restore market efficiency.
Collaboration in finance is largely uncommon and impractical due to the competitive edge that is characteristic to the kind of predictive modeling that is associated with the area. Numerai addresses this via running rounds of successive Machine Learning competitions for the purpose of crowdsourcing predictions, which then serve to form trading decisions on the traditional markets.
- A meta-model that combines crowd-sourced, pre-filtered proxies for trading decisions.
- Data anonymization and ensemble learning enable extremely difficult to outperform meta-model.
- A cleverly designed auction mechanism which crowdsources auxiliary metadata about the submitted models, namely the corresponding degree of confidence, by introducing clear financial incentive, via Numerai’s own ERC20 token.
As of the time of recording this episode, Numeraire (NMR) was listed on the following exchanges:
- Numerai website: https://numer.ai/
- Numerai whitepaper: https://numer.ai/whitepaper.pdf
- Numeraire (NMR) markets: https://coinmarketcap.com/currencies/numeraire/#markets
- Kaggle: https://www.kaggle.com/
- Log Loss metric: https://wiki.fast.ai/index.php/Log_Loss
- Numeraire ERC20 smart contract: https://github.com/numerai/contract
- Deep Learning, by Ian Goodfellow (Hardback copy): https://amzn.to/2Ie20Uf
- Deep Learning, by Ian Goodfellow (free online version): https://www.deeplearningbook.org/1
The figures from the example, presented in the episode, were as follows:
- Prize pool: $1000
bid = (c=5 NMR, s=1000 NMR),
p = $200,
t = 1,
d = 0 NMR,
w = $200
bid = (c=1 NMR, s=200 NMR),
p = $200,
t = -1,
d = 200 NMR,
w = $0
bid = (c=0.5 NMR, s=500 NMR),
p = $1000,
t = 1,
d = 0 NMR,
w = $800, where
t- model result,
d- tokens destroyed,
w- actual winnings, and
p = s/c- potential winnings.
The actual reward is calculated as
R[i] = min(p[i], r), where
i denotes the potential winnings of the
ith participant (ranked in descending order of confidence); and
r denotes the remaining prize pool amount.
The free electronic version of the book is, in fact, only available in HTML format, rather than PDF (the latter was originally indicated in the recording). ↩
PlatformsThis podcast can be obtained from one of the following platforms:
Please, consider providing an honest rating and review on your favorite podcasting platform, in order to facilitate ease of discovery by listeners with shared interests. This way, we can reach out to larger audience and achieve higher quality of constructive discussions and a healthy diversity of topics and opinions.
The tracks used in this episode have been licensed under the Creative Commons license. Please, find the corresponding artist, track names and links as follows:
- Opening track: KELLEE MAIZE - In Tune (J. Glaze Remix) (2014)
- Closing track: DISCONAUTS MUSIC PRODUCTION - Sunshine (2012)
- Section breaks: YUANAN "J.A.R" - Jammys comes to dinner Produced BY J.A.R 036 (2017)
Please, keep in mind that I am not a financial advisor and this is not financial advice. The sole purpose of this podcast is entertainment, and mutual education via constructive discussion.
The latter is encouraged via one of the available communication channels.