Thanks Dan,

I read the page, https://www.quantopian.com/fund (as an aside, gotta love your webpage overlaid on a blinding snowstorm!). It clears a few things up. Some feedback:

1. Presently, it sounds like participating in the contest is the only way to be considered for hedge fund inclusion, correct?
   
2. On https://www.quantopian.com/managers, you state clearly "Only strategies with 6 months of real money live trading track record on the Quantopian platform will be considered." This implies that the contest is the only way to be considered, one would have to win the contest, and after 6 months, a decision would be made.
   
3. I don't quite follow the statement "The individual algorithms don't need to be exceptionally profitable; the key to the enterprise is that the algorithms be uncorrelated." In the limit that all of the algorithms are unprofitable (i.e. consistently zero return), doesn't this whole premise fall apart? There must be some sweet spot, right? Or is there some financial wizardry that would take a bunch of uncorrelated strategies, all returning a few percent per year into a fund that would return 20% per year?
   
4. "We don't look at your algorithm code." Suppose I want you to look at my code? Or I decide to post it publicly? Could the algo still be in the hedge fund? If figure that the easiest way to convey "Strategic Intent" would be to reveal the code.
   
5. "Our target is to pay you 10% of the profit on your algorithm's allocation." What's preventing you from making a definite commitment ("Our target...")? Not so motivating. Say somebody spends the next 6 months, nights and weekends developing an algo, and it turns out you'll only pay 2% instead of 10%. How about "We will pay you at least 10% of the profit on your algorithm's allocation"? Note also that on https://www.quantopian.com/managers, you state "An industry competitive share of the performance fee; possibly as high as ~15% of the performance generated by the algorithm." Are you scaling back what you'd share?
   
6. "Financial theory, backed by evidence in practice, shows that by aggregating a large number of uncorrelated returns streams we can deliver overall predictable returns to our investors." You speak as if there are well-established examples of scalable strategies using real money. Do you have case studies to share? It sounds like you have a specific, well-established fund formula in mind. What is it?
   
7. You state "We are looking for algorithms with a market beta between -0.3 and +0.3" and then later, "We prefer that average to be between -30% and +30% average pairwise correlation with the other algorithms in our portfolio." So, it sounds like in addition to the low beta (which algo developers can compute), you are also requiring a low correlation to other algorithms. To develop an algorithm, this means entering it in the contest for evaluation, right? Or would you expect the low beta to ensure the low correlation with other algorithms? It just seems like a slow workflow to have to enter the contest to get feedback on the "low correlation to other algorithms" requirement.

Grant

I don't quite follow the statement "The individual algorithms don't need to be exceptionally profitable; the key to the enterprise is that the algorithms be uncorrelated." In the limit that all of the algorithms are unprofitable (i.e. consistently zero return), doesn't this whole premise fall apart? There must be some sweet spot, right? Or is there some financial wizardry that would take a bunch of uncorrelated strategies, all returning a few percent per year into a fund that would return 20% per year?

Love this.

"We don't look at your algorithm code." Suppose I want you to look at my code? Or I decide to post it publicly? Could the algo still be in the hedge fund? If figure that the easiest way to convey "Strategic Intent" would be to reveal the code.

Love this too. Grant, what if in the future, you start your own hedge fund, can you tell Q to stop using your algo?

"Our target is to pay you 10% of the profit on your algorithm's allocation." What's preventing you from making a definite commitment ("Our target...")? Not so motivating. Say somebody spends the next 6 months, nights and weekends developing an algo, and it turns out you'll only pay 2% instead of 10%. How about "We will pay you at least 10% of the profit on your algorithm's allocation"? Note also that on https://www.quantopian.com/managers, you state "An industry competitive share of the performance fee; possibly as high as ~15% of the performance generated by the algorithm." Are you scaling back what you'd share?

Love it! It is their lawyers talking "Our target is to pay you 10% of the profit." This will give them the opportunity to reduce the % in the future.

"Financial theory, backed by evidence in practice, shows that by aggregating a large number of uncorrelated returns streams we can deliver overall predictable returns to our investors." You speak as if there are well-established examples of scalable strategies using real money. Do you have case studies to share? It sounds like you have a specific, well-established fund formula in mind. What is it?

Grant, you are really speaking your mind.Be careful ... as they will find ways to return the favor in the future.

Grant,
Many good questions. I'd point out that "aggregating a large number of uncorrelated returns streams we can deliver overall predictable returns to our investors." comes straight out of CAPM, which has a huge body of research around it. Certainly room to disagree with the idea that they can put something together that reliably provides alpha, and even room to disagree with the CAPM concept, but the underlying theory is pretty well represented in the literature and practice so they're not doing anything revolutionary with the idea itself.

Grant,

Here are a couple of screenshots pulled from a couple of our Quantcon presentations. First screenshot shows how a diversified portfolio of uncorrelated assets significantly reduces overall volatility, and then the second screenshot takes it the next step of illustrating how this can be extended to increasing the Sharpe Ratio at the portfolio level. Specifically how a portfolio of pretty low Sharpe Ratio algos (algos with only a Sharpe Ratio of 0.2), when added to a portfolio, but with the stipulation that all the algos are uncorrelated, yield a portfolio Sharpe Ratio well over 1.0+. We can see about sharing an ipython notebook in our Research environment with this simulation if it would be helpful.

Grant first:
1) The contest is the front door to get in the hedge fund. I certainly won't rule out other ways to enter the contest in the future. I can imagine us creating other ways to submit algos, or maybe we'd pick up an algo writer through personal contacts at a Meetup, or something like that. But for now - the contest is the right way.
2) We've learned a lot since we first wrote the /managers page, and it became outdated. That quote is obsolete. Real money trading is not required to get selected for the hedge fund. I think we've been saying that for a while now, but we didn't correct that /managers page. Glad it's fixed now.
3) What we're saying is that we don't need a bunch of 10+ Sharpe algos, that more reasonable levels of profitability are sufficient, such as a Sharpe of 1 or higher. There is indeed some wizardry that you can combine uncorrelated returns stream into a higher overall performance - Justin covered that just above. In the limit that the algos are not profitable it does indeed fall apart.
4) It's your intellectual property, and you can share it with whomever you wish, including with us. That sentence is best read in the broader context of the full paragraph. It's not a prohibition for you. It's a commitment by us.
5) We're not willing to make a final and definite commitment to 10% for the exact same reason that the language is different from the older /manager's page: This is still a work in progress. Specifically, the old 15% number was in the context of low-capacity, high-Sharpe algos, and as you have read, we've modified the overall fund strategy. The other significant variable here is the size of the allocation. By going for these more moderate Sharpe strategies, we can find larger capacity, make larger allocations, and the actual take-home pay for the quant goes up.
6) Covered this one I think.
7) I agree that the low pairwise correlation workflow is slow, and you make a very good point. It's a hard problem to solve. I'll see if we can come up with smart proxies that have a faster feedback loop. As you suggest, low beta is one of those proxies.

Uncle Bob, honest and well-intentioned feedback is always welcome here.

Out of curiosity, does trading my strategy with real money increase my chances of getting chosen at all? I already have my algo in the Open. Does a real-money record make you guys more inclined to choose it?

We're considering algorithms entered in the contest and those trading with real money equally for the fund. Having one doesn't increase your chances over the other.

@ Justin & Dan, Thanks for the feedback. I'll have to digest this business of uncorrelated algos. It just seems like you are driving down the variability in the mean. Of course you'll have a nice, stable mean if you construct the portfolio with a lots of uncorrelated, but "noisy" portfolios. It is just averaging, right? I don't see how this bumps up the return, which is money in investor's pockets. If the Sharpe goes up, due to lower volatility, then big deal. Can't I get infinite Sharpe by going to my local bank and taking out a CD (since the denominator is zero)? You must be assuming that the return will be high because each algo will have a relatively high return? Or is there some "Wow!" I'm missing in your analysis that boosts the return?

@ Alisa, I thought Dan just said that the only way to get into the fund presently is to have an algo in the contest. Now you are saying that real-money algos are being considered, too?

we have decided to open up our fund selection process to algorithms with "paper" track records, as well as those meeting the more traditional requirement of "real money" track record.
source Mar 6, 2015

Does that refer only to contest entries?
Or will paper trading algos that users start in their own working environment also be considered for the fund?

Dan, I'm actually quite disappointed with this change. When I design algorithms, I'm trading for myself. This means I prioritize Alpha and Sharpe as the two most important criteria. Beta is important only in a peripheral sense; I have no absolute requirements for it. High beta will not keep my from deploying an otherwise profitable strategy.

What this now means is that the algorithms I tune for my own profitability will not perform well for the contest. This result seems misaligned both with my individual interest, and misaligned with Quantopian's interest as well.

I suspect other algo authors will feel similarly to me. I think this change is going to artificially restrict the number of candidate algorithms you will evaluate for your funds, and filter out some high performing results.

They need a fund that looks like a bank CD, except with higher return. This is what their investors want, I guess. So, to get something that goes like (1+r)^n, where r is the highly stable return, and n is the number of times compounded, they need uncorrelated algos. My hunch, though, is that if an investor showed up at Q with $100M and asked for a high-beta strategy, they wouldn't send him packing.

What about having 2 Quantopian funds, a 'Low Beta/Risk Off' fund and an 'Risk On' fund. The Risk On fund does not necessarily have to be high Beta, tied to the S&P but it would have higher volatility than the Low Beta fund. The combination of 2 funds would give Q a lot more flexibility to move AUM from one market environment to the other and back. Limiting selection of systems to just the 'Low Beta/Risk Off' seems very limiting.

Keep in mind that hedge funds aren't compensated like mutual funds and as a result have a different focus. If you run a mutual fund you're expected to simply beat whatever index you're working off, the S&P 500 if you're a general fund for example. This means you can have -10% returns in a year when the S&P 500 is down 20% and you're a rock star. Your comp is generally tied to only to AUM. In a hedge fund you're mostly compensated by a percentage of profits over a hurdle rate, regardless of what the rest of the market is doing. That same manager in a hedge fund gets nothing but their management fee if they don't beat their hurdle rate. Hence the word "hedge" in the name and the general focus on low beta strategies in the hedge fund world. As an investor you'd just go with a mutual fund or ETF if you were looking for a high beta strategy, plenty of those with lower fees. Always exceptions to the rule, but when you're trying to fundraise for a new concept in hedge funds with an inexperienced team you probably want to stick with the investment thesis that most potential LPs are looking for.

Grant Kiehne,
You absolutely right about CD like investment.
Not only Sharpe Ratio will get rank 1 as StDev =0,
Calmar Ratio will get rank 1 as Maximum Drowdown =0,
Volatility will get rank 1, as it will be zero,
Beta will get rank 1 as it will be zero,
Stability "of doing nothing" will be ranked in top ten,
Consistency "of doing nothing" will be near 1.
And it doesn't metter what is your return rank 100 or 300 because 6 not money related factors
already open the door to Quantopian Hedge Fund Management.
How they gonna make money?
Exposer to market is not only risk it is an opportunity to make money.

https://www.quantopian.com/posts/how-stable-is-stability-calculation
https://www.quantopian.com/posts/how-consistent-is-consistency-factor

Grant, Ken, [others asking about why we're not interested in high beta strategies],

One of the reasons why we're first targeting a "High Alpha, Low Beta" portfolio as our primary focus right now is by and large because that is one of the most marketable hedge fund products to institutional investors. Historically these types of strategies have been called "portable alpha" or "pure alpha" strategies. In fact, BGI (now Blackrock) pioneered the concept of portable alpha a couple of decades ago and ran quite a successful business doing so. Bridgewater's flagship strategy is called "Pure Alpha" if I recall... there are lots of successful case studies in this type of "High Alpha, Low Beta" portfolio being an appropriate manner in which to focus effort.

So why don't institutions care about "High Beta, High Alpha" strategies as much? Quite simply because Beta is pretty easy to get. Just decide how much Beta you want and go buy S&P Futures that require only about 5% margin requirement per unit of Beta that you need ( only $23,000 to get $500,000 S&P500 exposure, http://www.cmegroup.com/trading/equity-index/us-index/sandp-500_performance_bonds.html , http://www.cmegroup.com/trading/equity-index/us-index/sandp-500_contract_specifications.html). Then you take the rest of your investable capital and invest that in "Alpha strategies." This is more common practice that you might expect actually. I actually think even some mutual funds are allowed to achieve their simple S&P500 exposure in this manner nowadays as long as they don't exceed leverage beyond regulatory mandates, since its way cheaper to trade futures contracts than it is to transact in 500 individual securities...

So let's say given the above numbers regarding margin requirements, and you have $1 million to invest. Let's say you love beta but also think you have some stock picking alpha. Cool. So if you want to be 3x Beta, that means you want your $1 million dollars to fluctuate as if it were $3 million. Easy, go buy 6 S&P Futures contracts (each one control ~$500,000 as described above), and this only ties up (per the link to the CME current margin requirements), 6 * $23,000 = ~$150,000. Great, now you have $850,000 to invest in "Pure Alpha". If you're an institution you might go looking around at hedge funds to allocate this... perhaps a hedge fund that aims to deliver only alpha, like a BGI, or a Bridgewater.... or as we hope, Quantopian :) So now your portfolio is 3x Beta exposure + Alpha, and the beta was achieved in the lowest cost manner possible.

I've definitely glossed over some of the operational details in the above on how is done in practice, but I hope I provided a bit of the insight (at least as how I see it), as to why we're aiming for a "High Alpha, Low Beta" portfolio -- namely because its "portable" to quite a diverse set of institutional investors for them to integrate into their portfolios as they see fit. So, if we aim to deliver a low beta, and hopefully very uncorrelated return stream, to the institution to integrate into their own portfolio (which is also very likely constructed overall in a diversified manner) then their adding of our low correlation return stream (e.g. the 'Q' hedge fund) improves their overall diversification in much the same way as Q adding another uncorrelated return stream to our hedge fund portfolio.

Then the question for me becomes Dan, are Quantopian contributors allowed to invest in the fund? If so that's fine and I'll try to create low beta strategies but if not I'm better of developing for myself only as I can create balanced strategies that are aligned to my risk profile. As I'm time poor I cannot do both..... What is the intention?

Justin, what is the target range of returns you hope to get from each Low Beta Q return stream? Best case, most likely?

Thanks Justin,

I don't follow all of the fancy finance stuff, but I think I get the gist of what you are saying. One piece that is missing is how much return the Q fund needs to be competitive. When I hear ultra-low-volatility, market-neutral, arbitrage, etc., I'm thinking you want something that goes like (1+r)^n. Oh, that looks like a CD! So then I go to http://www.bankrate.com/ and see that my reference rate of return for a 1-year CD is about 1.3%. So, if I can write an algo that does better than that, I'll be golden.

I'm guessing that the Q fund won't be competitive if the return goal isn't something more ambitious. Is there a number you'd like to hit? And why? In my mind, talking about Sharpe without return doesn't really help. I think that you need to publish some sort of guidance. For example, if my low-beta strategy doesn't meet or beat the long-term return of the S&P 500, will it be attractive? Are there reference returns for the hedge funds you mention, or are they shrouded in secrecy?

Grant

@ Dan and "Quantopianians" :-)

First of all, many thanks for increasing the size of Tradeable Universe to 500 Securities!

Which timing/scheduling have you planned for the start up of the hedge fund?

Will there be regular (e.g. yearly) deadlines for entering/selection?
Or admission to the fund will always be possible, given the availability of slots and suitable algos?

Will a formal submission be required, like e.g. for entering in the contest?
Or are you planning to directly contact the coders of suitable, already running, algos?

What do you think about possible competition between an algo selected for the fund and the same algo continuing to trade with real money for its coder? Or between algos of your pool competing on the same Securities?
Could a "crowding the trade" risk arise?

It sounds like Q is going after sophisticated institutional investors with the Q Fund (i.e. Wall Street...I thought Q was anti-Wall Steet? I get so confused sometimes!). What about the retail market of "traders" (i.e. gamblers/speculators)? Wouldn't it make sense for Q to also open up a trading venue, so that Q algo writers could write automated trading vehicles for the retail market? There must be lots of people out there who'd want to throw money at a speculative, high-beta algo running on Quantopian, right? Or would there be regulatory problems with opening an online casino...uh, I meant trading venue?

Hi Grant - institutions are the customers of Wall Street. We want those customers. They are universities, pension funds, family offices, hospitals, and sovereign wealth funds. These are institutions that serve society, and we want to serve them directly. My view is that there is a dearth of hedge funds providing algorithmically managed funds to them.

The retail market depends on marketing at an incredible scale and driving costs to zero. Successful companies like Vanguard (low cost winner), Schwab, and Fidelity are competing with startups using internet marketing combined with very low cost products (WealthFront, Betterment, Robinhood). I admire those companies (new and old), but my dream is to connect institutional investors with independent quants.

@Dan, seems like it's more like "it's all about that alpha, 'bout that alpha - NO beta" :) or pick your desirable metric.

An interesting and positive modification to the contest for sure, depending upon the time interval over which you calculate the net beta. The longer backtest, sure -- but, real-time over the current month might be an issue since a low beta strategy over time can certainly go +/- a much larger value over a smaller time calculation interval. At least that's my thought of the evaluation - I couldn't find the methodology spelled out on the site or in the Help. (yes, i could dig in to the source :) )

The enormous Sharpes I see often on the leaderboard still puzzle me (overfit?) for success over time and walk forward, but that's an entirely different issue.

Anyway - glad to see the continued maturing of the product, a very interesting experiment indeed. Once futures are added to the mix, and if permitted in the contest (with a mandatory cash or STIR bond ratio reserve -- gotta be viable for real), this will be worth some extra effort from me on the platform proper vs. offline.

Great discussions as always!

Thanks Fawce,

I have to say you guys have a way to dribbling out what you're aiming for. I guess it is all obvious to folks in the industry that you'd go after "universities, pension funds, family offices, hospitals, and sovereign wealth funds." I figured you were lining up a bunch of 1% types not knowing what to do with all of their money (I guess that's covered by "family offices"). And if you aim to steal business away from the establishment, then I'll credit you with aiming to "Hack Wall Street."

I'm surprised that the institutions would all want low beta, absolute return strategies. Or I guess you are thinking that's there's just too much competition in the world of index tracking / beating instruments? Or it's not profitable enough?

I was thinking more along the lines of Wild West gambling-style trading, where somebody with speculative money who doesn't know how to code would be hooked up with a Q quant. A pure retail casino model. The only service to society would be entertainment, like buying a lottery ticket or going to Vegas. My hunch is that there could be a big market. And Q could provide the liability and legal support. I'm kinda joking, but my sense is that there is a whole market you are avoiding.

Grant

Hello Fawce,

Some questions:

1. For proof-of-concept (I'm gathering around $25M in capital), where's the money coming from? Your VC's? Or do you have a lead institution lined up? If so, who is it?
   
2. There's been a lot of concern/guidance on regulatory constraints on your Q Fund concept. I'm just trying to imagine someone in charge of investing at a university going to his administration/board and saying he'd like to invest $250M in a new hedge fund, composed of a bunch of black-box algos developed by random eggheads from around the globe, sitting in their pajamas coding up strategies. It seems like you'll end up with an unconventional product aimed at conventional, conservative, well-established markets that will beat you over the head with legal requirements and specific demands on how you'll need to run things under the hood.
   
3. Overall, I'd be concerned that you aren't aiming to make a market, but rather compete in an established one. As a start up, you are basically saying that you can overturn the incumbents at their own game, rather than create a new market. It's low risk to say to your VC's that you'll be able to access an established market. What will be your compelling story, when you stand up in front of the guy who runs a big pension fund that'll make him want to give the Q Fund a try?

Grant

Arthur C. Clarke's Three Laws:
1. When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong.
2. The only way of discovering the limits of the possible is to venture a little way past them into the impossible.
3. Any sufficiently advanced technology is indistinguishable from magic. :)

William Hutchison Murray ...
"Until one is committed, there is hesitancy, the chance to draw back, always ineffectiveness. Concerning all acts of initiative (and creation), there is one elementary truth that ignorance of which kills countless ideas and splendid plans: that the moment one definitely commits oneself, then Providence moves too. All sorts of things occur to help one that would never otherwise have occurred. A whole stream of events issues from the decision, raising in one's favor all manner of unforeseen incidents and meetings and material assistance, which no man could have dreamed would have come his way. Whatever you can do, or dream you can do, begin it. Boldness has genius, power, and magic in it. Begin it now."

"There's a sucker born every minute" is also a quote to keep in mind.

I must agree with Grant in that there is a huge market for hedge fund like strategies executed for people who do not have access to hedge funds, or more reliable return streams. I'll be happy to invest in the hedge fund myself wether my algo gets selected or not. And talking to my friends they would love to. They even offered to co-ride my algo which I friendly refused as I don't want to ruin relationships. However if the algo would be in the fund I would recommend them to invest in the fund as well.

On a different note I still don't understand the sudden focus on beta. I managed to combine 2 of my Algos One with with high negative beta and one with high positive beta. So now the outcome is beta-low..... But now I created a less pure strategy that adheres to the goal of low beta but I would rather invest in both Algos separately and let them show the pure performance. I could make the mix more optimised with an allocation algorithm that dials one algo up and the other down to maximise returns but I thought that would be the role of the black box hedge fund..... I wonder whether the low beta requirement will create little hedge fund Algos that will not have a lot of tuning opportunities for the black box.... Or... Maybe I just don't understand this area enough

Fawce (if I may call you that) - targeting the Wealthfront type customers with a more advanced product seems like the obvious move here!

If you can land institutional investors, that's great! You've suddenly got a very large bankroll! But that's a big 'if' you'll be able to answer soon enough.

Wealthfront customer persona are hands off investors that just want good performance over time with low fees.

If you fit into that space of net worth of 1-3M committing to a hedge fund isn't possible:
1. you'd be putting all your eggs in one basket.
2. you have liquidity needs like the possibility of purchasing real estate

That leaves you with:
1. investment advisors - which leaves you with an oily 'sales-y' type feeling that they just want to skim money in fees off the top. I'm talking about Fisher Investments, Chase Private Client, MSSB, or just general 'advisor' firms. They want to sell you some products for 100K-500K min and charge 2% management with no performance hurdles. When the market tanks, you're going with it, and they're still taking their 2%.

1. investing along for the ride with someone like Warren Buffett through BRK or investing in what you know (hurray for Apple) - a little scary when you're starting out investing serious amounts of your net worth.

2. something like Wealthfront which at least you feel like you're going to to 'keep up' with the markets and you're not getting ripped off by fees. if the market tanks, you'll tank, but at least you know everyone else did too and they didn't take outrageous fees (0.25%)

3. buy into a smaller fund that promises hedging strategies with low overhead costs, perhaps with lower minimums than a typical hedge fund. This type of product doesn't really exist for automated trading strategies.

Props for a cool product and great engagement from people in the community! Don't know how they keep up with all the action and still have full time jobs ;)

I want to keep this short and sweet as I'm short on time and have a flight early tomorrow and I want to be well rested for traveling.

I have been coding on Quantopian for a few months now steadily improving in knowledge and skill. I had come up with a extremely profitable algo for this months contest only to find out about the new Beta restrictions. I agree with what was stated earlier by Ken to me personally the most important measure seems to be without a doubt the Sharpe ratio.

I'm just struggling to understand why the sudden Beta change was made. I still dream of running my own algo on a large account of my own in which Beta is irrelevant the only thing that matters is profitability and like Ken mentioned earlier now that mine and Q's priorities seemed to have fundamentally shifted. This is putting me in a weird state of mind because on one hand I have enjoyed participating in the Open, but at the same time that was just a side effect of developing my own algo to run on my own account.

I spent a bit of time trying to get a algo that can run within the stipulations forced on me by the new rules and I came up with something. The thing is I not really sure what it is as the Sharpe ratio was severely hamstrung compared to previous algos I had tested without the lower Beta requirement. Long story short I and seriously considering not taking part in the Open next month if the Beta rule still applies. I will still use the tools and resources available on Q, but really only for developing my own personal algos.

I was wondering if any of Q's admins could shed some light on whether the Beta rule will still apply in the July Contest? I haven't seen anything about July's contest mentioned yet.

Just voicing my concerns sorry if it felt like I was rambling.

Sincerely,
Spencer Singleton

As Justin says above,

One of the reasons why we're first targeting a "High Alpha, Low Beta" portfolio as our primary focus right now is by and large because that is one of the most marketable hedge fund products to institutional investors.

I gather that Q is pulling a "me too" play here, in that there is an established, large, and potentially profitable market for high alpha/low beta hedge fund products. Their VC's don't want to hear about creating new markets; they'd settle for a new category within an established market. To pull this off, they need to drive Q users to look for money left on the table, to figure out how to print money. The business about correlation of returns, in my mind, is irrelevant. They really want every algo to behave like a bank CD, with high return. In this case, the returns will be highly correlated--they'll all march consistently upward! If every algo consistently returned 20% per year, in a (1+r)^n fashion, I don't think Q would be disappointed. They are just saying that if there is variability in returns, it would be nice if it were uncorrelated, so that the overall fund variability could be driven down, by averaging (this is the no-brainer part of the story).

A couple technical questions on the beta calculation:
1. Are you keeping the Beta score component, i.e. there are now 2 Beta factors, a score and a screen?
2. You describe a 12 month rolling annual average Beta calculation. Is this used for the score and the screen Beta, or just the score?
3. Is this rolling average Beta used even with an algo that has more than 2 months of paper trading? Right now if you have more than 2 months of paper trading the backtest ceases to be used at all, from what I understand. It appears that backtest results will persist in some form for at least 12 months going forward?
4. If 3 is correct, what happens with strategies that use newly introduced securities? Right now you get a hit in the first month because of the backtest, but by the second month you can compete on a level playing field with securities that were introduced in the last 2 months. If 3 is correct, you essentially won't be able to use a strategy unless its component securities have been trading for over a year without a significant performance hit?

Justin,

I had a question, and some things to think about, regarding correlation.

How can we make uncorrelated bets in the hedge fund, or any algos, if, as far as I am aware, Quantopian only allows trading in stocks. Stocks typically carry a 0.6 correlation to other stocks. I am a big believer in the power of uncorrelated bets. I think it is absolutely the right move, and have heard it said by one prominent investor that the holy grail of investing is to find 15 uncorrelated bets, and that is because it can reduce your risk/return ratio by a factor of 5.

The trick is finding those uncorrelated bets, and from what I have experienced, you have to go outside the stock markets to find those, and into things like spread positions, i.e. bond spreads, currency spreads, etc. Can you do that on Quantopian?

Also, one thing to discuss/keep in mind, is that correlation doesn’t really exist the way people think it does. That may sound strange, but correlation is just a word people use when they see two markets moving in tandem. Gold prices and the dollar rise and fall in opposite directions, so they are said to be negatively correlated. That is wrong. What is really happening is that individual markets are behaving in a way that is logical based on the underlying drivers of that individual market.

So, the key becomes finding the cause of the correlation…the underlying drivers of the individual markets.
That is where you can generate Alpha.

Grant,

I like your idea of offering high beta strategies to the retail market. With the recent move toward robo advising (only allocating across three asset classes) it only seems natural that hedge funds would follow. The problem (i believe) is that investors in a hedge fund are required to be qualified investors (high net worth). There is some rule that you can have 20 or 30 non-qual investors but that won't get you very far. If anyone knows a way to offer retail clients hedge fund like investments.. let me know--I've been trying to think of ways to do this.

Jamie,

I think you are referring to my "pure retail casino model" comment above. I wasn't thinking about a hedge fund, but rather that people like to gamble, and the stock market is one place that they do it. Quantopian algo writers could supply risky strategies for the retail gambling market, and Quantopian could take a cut, while also providing legal and liability coverage. No hedging. Just dice rolling. But I'm guessing it'd be illegal (and kinda sleazy, too).

Grant

Let's open a Kickstarter campaign where we'll take contributions at $1000 a pop down to, oh, say $100. We're going to build a trading software product that we'll share after "development" is complete and it has been "productized". Everyone will get a copy of the software. Oh, and we'll send everyone a nifty Kelly (Criterion) green tee-shirt and take the money and invest it into "the product." In the mean time, to ensure everyone's best interests are at heart, we'll invest the whole she-bang in an IB account and trade it using the best 5 algos (as governed by a magical metric manifested from the measurements and methods derived from the market molecules). At the end of say, 18 months we call the software a failure and return everybody's donations, plus whatever they happened to have earned during the development and minus some small fee for the tee-shirts. And we also hand over the software as consolation for the time we wasted with their contributions tied up.

The success ratio of Kickstarter and Indiegogo must be like, what, 10%? This will have a success ratio way better than that.

So, who's a lawyer here? What are the implications of doing this? And who's going to jail first?

My other thoughts about this discussion is that this isn't so much about "not beta" but about generating uncorrelated return streams - if one uses low beta as a indicator for non-correlation, not unreasonable as a filter Q has now employed. Lisa Borland, at QuantCon, noted the different "kinds" of algos there appeared to be in action from observing the results, with particular attention paid to those in one corner of her chart that were considered uncorrelated data streams (wrt the market).

I'm not saying that this (only) is the way to run a portfolio, although some funds certainly do; but this combined with a traditional beta portion could certainly be a nice way to improve overall performance (as noted early on in the meaty textbook, Trend Following with Managed Futures: The Search for Crisis Alpha by Alex Greyserman & Kathryn Kaminski). Looking at an appropriate non-beta algo over time in the platform might not appear as appealing on first inspection, but in beta combo (for portfolio not contest purposes)... HTH.

Back to the current topic in this thread, looking forward for Grant & MT to open up shop :).

To pick up on Philip's point,

The trick is finding those uncorrelated bets, and from what I have experienced, you have to go outside the stock markets to find those...

It would be helpful to have 5-10 distinct example algos developed by the Quantopian team that they would consider for the crowd-sourced Q fund. The algos would serve as an existence proof that there is a wealth of opportunities for the crowd. Just replace the example algos on https://www.quantopian.com/help with relevant ones. In the end, to get to $10B, at least 400 algos will be needed, if not thousands (since each will get only $25M max), so if 5-10 can't be conjured up quickly by relative experts, then where does that leave the crowd?

It would be helpful to have 5-10 distinct example algos developed by the Quantopian team that they would consider for the crowd-sourced Q fund. The algos would serve as an existence proof that there is a wealth of opportunities for the crowd. . . . if 5-10 can't be conjured up quickly by relative experts, then where does that leave the crowd?

My feeling is that producing example algorithms would prejudice the pool; the point of crowd-sourcing is to get the crowd searching widely across the space of possible algorithms, constrained by the rules and motivated by the value function. If they toss out 5-10 example algos, especially with the implication that these would be fund candidates, a large percentage of the agents (quants) will switch from wherever they were searching to investigating variants of the examples. Although perhaps cruel, if 99 quants search futilely in neglected parts of the search space, but one quant finds something truly novel, that would seem to be a better outcome for Quantopian than 100 quants all searching in the well-trodden and intercorrelated space around the example clusters.

As for low beta algos not being possible with only equities, that's simply false, you just need to be open to shorting things. In fact, given how the math works, if your goal is to shoot for a fund allocation with or without winning the contest, you'd probably be better off aiming for negative SPY beta; even if the returns are mediocre, if the algo consistently does well when the market is doing poorly, that's gold (figuratively).

As an aside, I expect there will be a first-mover advantage to those folks who start entering futures-based algorithms into the contest as soon as it is possible to do so.

Hi Simon,

Read your reply, and I agree with most of what you are saying. Some thoughts: At the end of the day, if the goal is to create a portfolio of truly uncorrelated assets, and I think it is, has anyone considered that correlation might be a poor way of diversifying away the risk? I know, it sounds a little crazy, but as Ray Dalio says, "Drivers are the cause, Correlations are the consequence."

Essentially, the thinking goes that correlations are backward looking and variable, and thus only useful, until they invariable fail, and are no longer useful. The real question is what is the cause of that correlation in the first place. Why do certain stocks move in tandem? What is the cause, the underlying driver? Figure out the underlying drivers moving a particular position, and then structure your portfolio so that the individual positions have individual drivers. At least, that's my basic understanding of how Bridgewater supposedly attacks the problem of diversifying risk away.

Would be fun to get a collaboration going, trying to figure out underlying drivers and how they can be used to create a truly diversified portfolio. Anyone interested?

True. No doubt that while for the moment, they are trying to isolate alpha with respect to CAPM, in the future as they get more institutional, they'll try to isolate some alpha with respect to a more complete model like Fama-French 3-factor or 5-factor, wherein those known drivers can be accounted for.

If you can come up with a screen or algo which is neutral to known drivers (or even just principal components), yet has residual alpha, bob's your uncle, you're done.

@ Simon,

Yeah, I thought of the fact that in providing actual hedge-fund-ready algos, Q potentially would be biasing their pool. However, if I grasp this whole crowd-sourced concept, they need on the order of 1000 unique algos to get to $10B, so 5-10 wouldn't make a difference. And Q would see the correlations and optimize out those example algos anyway (maybe just keeping the original 5-10). Besides, eventually one would hope lots of viable examples get posted here. And if I understand Dan correctly, since managers would own their algo IP, they could share their algos or even post them publicly or perhaps sell them to other Quantopian users to tweak and submit for consideration. There would be no protection against the problem of users gravitating toward a limited set of strategies.

Regarding your statement "As for low beta algos not being possible with only equities, that's simply false, you just need to be open to shorting things," it may well be true. If you care to prove it, post some algorithms! : )

Well, it's almost true by definition - short SPY and you will have a CAPM beta of -1...

I think they are looking for beta around zero, and high alpha, right? Do you think that there are 1000 unique pure alpha strategies to be had using Quantopian, even if all 35,000 users really sharpen their pencils? That's the question, right?

Perhaps, perhaps not, but the solution is not to relax beta constraints; people don't pay performance fees for beta, if they want beta they can get it for like 3bp.

It brings up an important point though, that if people believe in the vision and want to participate in the fund, they'd be wise to start thinking about capacity now. Strategies that are great for personal accounts or for the Open might not be so scalable when trying to capture performance fees on $100M+.

A topic for a separate thread, but I wonder how the folks at Q plan to handle the scaling? Maybe they figure the slippage model will be good enough up to $25M to get a sense for how much capital a given strategy will support? And then test the waters by gradually ramping up the capital? And I figure they look at the securities that are being traded, which would help understand the capacity (for the dashboard of my winning algo, they are visible, and Q must have access to the IB account, as well, which I cannot see).

Simon - Great points, thanks for the thoughtful posts! While I agree wholeheartedly that novel strategies (and lots of them!) are what we're all after here, I do think some more basic examples of how to build a market neutral strategy like Grant's asking for would be helpful to some of the community who are just getting started.

Here's one very simple one I've been playing around with as a teaching example and would love feedback on.

It's a market neutral 5-day price reversal strategy. Every week, you rank all the stocks in your universe based on the prior 5-day returns, buy an equal weighted basket of the 10% of stocks with the biggest losses, sell short an equal weighted basket of the 10% of stocks with the biggest gains.

What's nice is that with the recent performance improvements that the engineering team's been rolling out, I can now run this over a reasonably large universe, the top 5% of stocks by trailing ADV using set_universe. This means that I'm trading a portfolio of about 40 longs and 40 shorts.

While the un-levered returns of a strategy like this might not look too impressive. If you do find a market neutral strategy that has very consistent profitability, now you can apply leverage to reach a higher annual returns target. For example, here are results for the same algo as above, now raising the long and short side leverage from 0.5 and -0.5 to 1.5 and -1.5 respectively.

Thanks Jess,

You might try a linear fit to the minute-level price data over a one week span, versus the gross percent change calculation (and figure out how best to incorporate all of the OHLC values). Then you could set thresholds on the goodness of fit and the slope, to determine true monotonic gainers versus losers. This would also have the effect of avoiding the noise in tic level data that Ernie Chan discussed at QuantCon (as I understand, Q is just sampling individual trades, just prior to the whole minute, to construct the minute bars).

By the way, why don't you have Thomas W. noodle on an algo that would work in the Q hedge fund and post it. If he can't come up with something, bright young man that he is, there's no hope for the rest of us!

Grant

Jessica,

First, thanks for posting. As a low-skilled programmer, I accomplish most of my work here by modifying other people's shared code.

I have been looking for an example to count and record the number of open and short positions. I see you used "context.portfolio.positions.itervalues():" but that is not documented anywhere that I could find. Can you help me find it?

UPDATE: Oops, looks like it is a standard Python thing, not a Quantopian specific function.
http://python-reference.readthedocs.org/en/latest/docs/dict/itervalues.html

Tristan

Thanks Tristan, Glad you found it useful!

Look like you found a reference for itervalues, I'd also suggest checking out the Python docs directly for some more context and examples! The reason you can use this built-in python operation is that the context.portfolio.positions object returns a standard python dictionary.

Best, Jess

@Dan. I understand that you are looking for strategies with beta within -0.3 an 0.3. Those that qualifies, you place a blue badge. You also mentioned that we can calculate our own beta via research platform. Can you illustrate that with Jessica's strategy as an example? I do have some algo that on minutely level, beta is within that range. However, on a monthly level it is outside that range. I can't verify that at the moment. Not sure how to do that in research. I'm guessing there're two ways to calculate beta on this platform - use backtest data and plug into research somehow and the other way is to duplicate the algo in research and calculate beta that way. Appreciate it if you can illustrate how this is done.

EDIT: Got it sorted. Managed to extract the data from backtest. Found it in the Risk Metrics Section - Returns tab and Benchmark tab. Calculated the beta that way. R-square super low though 0.01.

Quantopian open June 2015 Average metrics of top 10 by stability (of loosing) and consistency (of doing nothing)

Quantopian open June 2015   annRet   annVol   maxDD   sharpe    calmar  stability  consistency  
Stability Best 10_pt       -130.52%  10.64%  -31.62%  -14.226   -4.835  0.973      0.807  
Stability Best 10_bt        -36.15%  14.27%  -73.24%   -3.590   -0.493  0.876  
Consistency Best 10_pt        1.31%  12.74%   -6.44%   -0.258   -0.194  0.128      0.962  
Consistency Best 10_bt       38.59%  15.54%  -13.62%    2.175    3.325  0.780  

Why Quantopian open June 2015 winner still not officially announced?

They don't want to announce a 'winner' until they are sure the recent rule changes will produce a real (re: profits) winner. Announcing 'winners' in a hedge fund contest that are losers does not help the narrative. Perhaps they should announce that the June contest was canceled because the rules changed right before the contest deadline. If they announce a winner using the new rules; the algo returns are probably an accident. If they use the old rules the algo is probably underwater profit wise. The current contest may be a little embarrassing....

Michael Van Kleeck is the winner of the June contest which Dan had shared in this thread, and the official announcement was here. Take a look at those code suggestions as you keep developing your algos!

Digging into your June entries, it appears that you stopped them early in the morning on July 1 and then resubmitted 3 new entries to the July contest. Since you stopped the algos, they were not read onto the leaderboard. This has been updated, to include all algos live as of June 30 at 4:00PM and you can see their final placement here.

Last month's Low Beta is next month's Highly Correlated.

For 7 months these guys have had ZERO returns. And that's low beta good?

High Anxiety? Low Beta!
If investors had their own Holy Grail, it might be a portfolio balanced at the so-called efficient frontier - where the highest expected return comes from the lowest expected risk. Investors tend to define risk based on where they see the greatest potential for losses.

One popular approach for mitigating losses concerns what Wall Street calls beta, meaning correlation with a market benchmark. When beta exceeds one, that means a stock's movement amplifies the performance of the market benchmark. A beta between zero and one indicates less dramatic movement than the benchmark, and a beta of zero means a stock moves in a way that has little to no correlation with the benchmark. Stocks with low beta tend to have more stability and less exposure to the overall mood of the market. These equities are in industries that have stable revenues that are minimally impacted by economic cycles. Things you need year round, like food and utilities, tend to come from companies whose stock have low beta.

Low beta investments tend to gain popularity whenever conditions become choppy and during selloffs. It just so happens that we are beginning to see some of that in the way U.S. markets are reacting to Greece's economic crisis, China's market downturn, and the rout in oil and gas prices. That led us to make our Low Beta motif this week's Motif of the Week.

https://www.motifinvesting.com/motifs/low-beta

About Beta calculation:

I've been playing with beta and came accross this old post. Dan, when you say "In practice, that means computing your beta to SPY for the year previous to April 30, 2015, for the year previous to March 31, 2015, for the year previous to February 28, 2015, etc. until we have 12 computations, and then averaging them.", more specifically "computing your beta to SPY for the year previous to April 30, 2015", what would the period be on that year beta calculation? I suppose you'd take the monthly return from April 2014, May 2014, [...], March 2015 and April 2015 and put them in the beta formula. Then again, you could also use daily returns, quarterly returns, etc...

In an effort to keep the beta of my strategies as close to 0 as possible I need to have more longs than shorts due to the increased volatility (and beta) of my short positions compared to my longs. Therefore my net exposure must be positive even though the beta stays consistently around 0.

I also like having a bond rotation strategy to round out my algorithms. But these assets have a negative beta on their own so I am long on them - further increasing my net exposure. I tend to have 60% of the portfolio long in individual stocks, 25% short individual stocks, and 15% long in a bond ETF (dynamically selecting this ETF based on market conditions).

I don't think this is a problem for the contest; but in order to be considered for an allocation is this a non-starter? Do you have any guidance on the maximum net exposure you'd consider? Also, would there be no interest in algorithms that have a small bond component to help reduce volatility and keep a near-zero beta?

Hello Stephen,

Great point, even with dollar neutrality there may be more beta exposure in your longs than your shorts or vice versa. There are multiple ways of handling this, I'll list them in rough priority order. However in your case you have a particularly extreme imbalance in beta, so it may be hard to solve. Any solution that attempts to hedge for beta bears 'estimation risk'. Basically you're making your model more complicated and giving it another failure point.

• Determining the structural reason behind the asymmetry and attempting to correct for it at the alpha computation level. This is preferred because attempts to correct after your alpha have been computed stand a good chance of mangling your alpha and reducing Sharpe. This means trying to figure out why one side has higher beta exposure, and finding some way to counteract that. One approach might be adding a volatility factor to your model to try to offset what you're seeing. Your want your alphas to be constructed with risk constraints in mind, so that the portfolio optimizer doesn't actually make that many changes.
  
• Adding another factor that is asymmetric in the other direction. Building new uncorrelated models will always help improve your overall predictions, a great way to do this is use alternative data.
  
• Buying into other assets or even classes (futures), in a way that helps your exposure balance out. This can be preferable because it can help diversify your return stream and reduce vol/risk.
  
• Determining the rough exposure mismatch and adjusting dollars long/short to account. This tends to run into issues because in order to meet beta risk constraints you lose dollar neutrality and vice versa. In order to receive an allocation from Quantopian an algorithm needs to be close to zero in dollar and beta exposures, any deviation will result in a smaller allocation. So whereas some dollar exposure is not a deal breaker, it needs to be small and it will limit your max allocation size.

One very strong constraint we want to see enforced by any algorithm that receives an allocation is that of position concentration size. We cannot have a large amount of cash be invested in any single asset. So the strategy of using a bond ETF would likely be a deal breaker. However, using futures might work.

See this algorithm for an example of using the optimize API: https://www.quantopian.com/posts/wsj-example-algorithm

The TL:DR is I recommend reviewing your alpha model and figuring out ways you can add more uncorrelated models that will help neutralize your exposures or at least make them less extreme, while also improving overall prediction quality.

I've read all of the above and am confused even more then before.
In order to maximize Alpha, we need to sacrifice Beta, Right? For example, when SPX goes down, we want to have portfolio biased toward negative Beta, thus increasing odds of maximizing Alpha. On opposite side: when SPX goes up, we want as higher correlation to SPX and we want to maximize Beta, thus increasing odds of maximizing Alpha. So, if someone writes an algo that maintains for example consistent 10xSPX return independent of SPX direction, but having huge Beta swings much higher then -0.3...0.3 contests threshold that would be not a winner? Something fishy is going on here.
On another note: Quantopian Beta calculations are based on correlation to SPX. I don't understand why they are not yet properly exposed to backtests, but my question is different? Why are they selected to be correlated to SPX? Don't you think that having current SPX history (which is long term up biased) introduces obvious bias in our algorithm construction/behavior? As implemented now, long term backtest with 0 Beta already biased toward positive returns. If you are so much concern with minimizing Beta (I mean Q management), shouldn't you be kosher in selecting right indicators?
Why not introduce random noisy signal with comparable to SPX frequency/amplitude, but long term neutral? It can be even random. You might even simulate long term up/down bias with it.
Otherwise, in my humble opinion, it just doesn't make sense to participate in contests, where technically purpose of those contests are not clearly defined. If the purpose is just to minimize abs(Beta), then it is probably not the right contest for me. If the purpose is to extract some Alpha, while maintaining minimized abs(Beta), then someone from Q should declare what targeted Alpha they are looking for...

Hello Igor,

Alpha and beta are completely independent quantities by definition. Betas are known, and therefore cheap to acquire, exposures. Things like market exposure, small cap exposure, etc. Alphas are defined as new return streams independent of known exposures. So if your algorithm returns are fed into a linear regression as the dependent variable, with known betas as independent variables, the higher the alpha value the more of your returns are new.

The most common beta and what we will refer to here as 'beta' is exposure to market movements. What you are describing is a market timing strategy, which tries to time the market and control beta exposure based on whether the strategy believes the market will go up and down. This is not truly controlling beta. Truly controlling beta means having a rolling beta exposure that rarely deviates from zero. AKA your returns just have nothing to do with the market at all. You don't care if it's going up, down, sideways, it's a different thing.

The point of the contest is to try to develop a strategy whose returns are based on alpha. Focusing on developing models which take advantage of new and previously unknown effects. You want a strategy whose returns, volatility, and other metrics will remain consistent out of sample when presented with new data. Alpha can come from anywhere, as long as it isn't correlated with known betas. If you're looking for ideas for generating alphas, I recommend checking out our partner data sets. https://www.quantopian.com/data

These lectures may be helpful:
https://www.quantopian.com/lectures#Long-Short-Equity
https://www.quantopian.com/lectures#Factor-Risk-Exposure
https://www.quantopian.com/lectures#Factor-Analysis

Delaney,
Thanks for attempt to answer. However, this is all just words.
If "The point of the contest is to try to develop a strategy whose returns are based on alpha" -- then you have to specify what particular Alpha you are looking for with above Beta restrictions. Why this is so difficult? Maybe DD should be restricted too? Then you have to specify that restriction. Otherwise, it's just unknown for developer what he is trying to achieve. Somehow, you figured out restrictions on Betas, as well as you came up with 3 leverage. Why is it so difficult to specify other parameters?
On the other hand: If Beta is so important, why yours backtests environment shows "wrong" Beta? Or maybe it is not wrong? Could you please provide the code for Beta calculations during backtests. If its "wrong", why this is so difficult to fix?
Also, at the top of the thread Dan states that: "until we have 12 computations, and then averaging them". What do you mean "averaging them"? So, if i have one month Beta at +100 and another month Beta at -100 that would be a winner?
What I mean again: if we want to be technical -- we should be precise in our language and implementations. Otherwise, its just talk...

Delaney,

It seems that I'll have to try and find another factor that has more of a higher beta component to it to offset the low beta of my current factors. A question on this though, won't most fundamental based factors have this similar issue? Companies with good fundamentals tend to be companies that exhibit lower risk metrics (more stable, less volatility etc.) so this mismatch will always be present if not paired with technical factors and/or alternate datasets as you suggest.

I'll look to drop the bond component although I feel like it makes intuitive sense to include bonds and is a prudent strategy. To shed some more light on what I'm doing, it allocates a small percentage to either short or long term treasuries (looks at momentum signals between the two) and adjusts the overall bond component based on overall market trends/momentum. When it's in short term treasuries, wouldn't this be preferred over just holding cash? I guess Quantopian would prefer to be neither cash or bonds, just always fully market neutral and 0 net leverage.

If net leverage is so important, why is it not included as a metric in the contest? This suggestion would hurt my own algorithms which are doing quite well; but I feel like it would better serve Quantopian in your effort to filter out allocation candidates.

Lastly, this is another example of why Quantopian-as-a-Marketplace makes sense for Quantopian to consider offering. These strategies I've developed (and similar ones other community members are working on) are effective at reducing risk and are based in sound investing reasoning - but they don't meet the requirements of Quantopian's market neutral fund. Market neutral strategies only represent about ~10% of the investment universe, I would think Quantopian would be interested in tapping into the rest of the investment market, and meet the needs of both institutional and retail investors with different investment requirements.