Hi Pravin -

You raise a good question. The other angle is how effective such a strategy would be. Intuitively, it is hard to imagine 3000 distinct factors, but maybe they sorta bubble up as uncorrelated transiently and with the right combination algo, it works? Or it is just a big marketing gimmick?

Hi Grant,

After talking to people in the industry, many of them have hundreds of factors to chose from. They apply some custom algos on these factors to find uncorrelated alphas and combine them.

The trick in this industry is new data streams apparently. We are competing with professionals who have access to data such Reuters starmine, Factset etc. and custom risk models such as Barra. With quantopian we are left to build everything ourselves and most data is premium and comes with a cost.

HI Pravin,
3000 alphas does not surprise me.
There is public evidence that WorldQuant and other firms generate these alphas with a machine. (The WC founder published a book, I can point to three interviews where people managing over $1B discussed using machines to generate alpha signal. )

As these large quant funds have access to similar tools for automating the alpha creation process, the differentiator is alternative data. You listed some sources there, but the big funds want data that no one else has. They have teams with multiple people trying to find new data sources. The workflow is something like: 1. get data, clean attach ticker, etc. 2. Generate potential alpha signals 3. look at Information Ratio of alpha signals using something like Alpha Lens 4. build real strategy (AlphaLens does not consider trading costs, slippage, etc.) 5. Evaluate out of sample

Quantopian provides a lot of stuff for free, so do the best with what you have. If you want to trade other things, and other data sources you can use Zipline the open source tool that powers Quantopian.

Hi Peter,

Thanks for the information. Do you have the name of the book?

I agree Quantopian offers a lot of stuff for free but I can always hope for more :). After all it is a symbiotic relationship. The better data and models we have access to the more algorithms for Quantopian to trade.

Best regards
Pravin

"Finding Alphas" by Igor Tulchinsky.
When I searched for this book on Amazon on my phone the search results were a bunch of romance novels about finding alpha males, it was very funny.

With quantopian we are left to build everything ourselves and most data is premium and comes with a cost.

Note that it is still possible to get an allocation by using the free portion of the data to develop an algo; it will be evaluated using the full data set. I don't know what happens if the algo doesn't get an allocation, but moves to an evaluation phase within the fund. Perhaps if it shows promise, the author gets to develop it further on Quantopian's dime, with access to the full data set?

I do think that Q could look at some bulk number crunching and factor generation that users with limited resources are incapable of performing. In some sense, they've outsourced this with the Alpha Vertex Precog data sets, which presumably could not be generated by a Q user on the Q platform.

Hi Grant -

As I understand the free portion of premium data sets are only available for certain period. Are you sure we can use them to develop an algo for allocation? Because I won't be able to use it for latest days.

Best regards,
Pravin

Yes. See https://www.quantopian.com/posts/getting-an-allocation-june-2017-update :

Every data set offered on Quantopian comes with a substantial amount of free sample data - if you build a great algo using free sample data our selection process will identify it, and we will validate the out of sample performance on our internal infrastructure.

I just don't know how things work if the algo is identified as decent. Presumably, the owner would then get free access to the data (exclusively for use on fund algos, I'd guess).

Thanks Grant. I never knew that. They should publicize this more. Let me play around with all the premium data free samples now.

Best regards,
Pravin

I didn't know that either, thought you had to get a top rank in the contest to receive an allocation. If backtests can result in an allocation as well, that makes the free sample data much more useful.

As for automated alpha generation, I read somewhere that the process generates alphas that are very weak so it's necessary to combine several of them to create a useful algorithm.