Hi Jared and QC-Team,
I want to share my opinion regarding the Alpha Market with you. First, I would like to emphasize that I like the idea of the Alpha Market and I hope it will attract more investors and quants in the future. But there are also a few things that bother me. To be constructive with my criticism I will propose possible solutions as well.
1. Submission criteria encourage developers to overfitting
In this I refer to the two performance criteria “PSR>80%” and “quick drawdown recovery”. I understand why they are included, but I fear it could have the opposite effect. The concern arose from my own experience.
I created several good strategies with Sharpe Ratio > 1, low volatility, low beta, small drawdowns etc. and most importantly, I didn’t try to optimize/fit to historical data and needed only a few weak assumptions, so I was confident they would produce similar results in the future. However, it often just missed at least one of the two requirements, so I started adjusting the algorithm. After a while, it met all the criteria for Alpha Market, but I was no longer satisfied with the result. I realized that I was overfitting the strategy.
I then decided against submitting the alpha. However, I doubt that everyone is that self-critical and that makes me doubt the quality of the alphas in the market.
Long story short, the two performance criteria “PSR > 80%” and “quick drawdown recovery” force many developers to overfitting, which in turn misses the point of the whole thing.
Suggestions for 1.: I would relax the performance-related criteria a bit and, in return, introduce other criteria that, as an overall package, ensure that only interesting alphas enter the market. More on this later.
2. Brokerage model, margin requirements and regulatory affairs
Currently, the Alpha Streams Brokerage Model is mandatory. This only allows a leverage of 1 for equities, i.e. trading on margin is not possible.
This made sense in the earlier version of the Alpha Market, where it was exclusively designed for professional investors. These then consumed the insights. But in the current version of the alpha market, it makes less sense in my view, because here the capital allocated for the alpha is managed entirely by the algorithm, including position size/ portfolio weights.
I think it would be better to use a brokerage model that is close to the broker used for live trading.
For equities this would typically be 50% as margin requirement (or leverage = 2).
Incidentally, currency pairs and futures can also be traded with margin, so that equities are treated differently here. The reason for this is not clear to me.
About regulatory affairs, just two examples:
Most US-ETFs are not tradable for retail traders from Europe. However, many Alphas trade these ETFs (SPY, TLT etc.), so there are errors in live trading.
Another examples is the pattern day trading rule (PDT) for small accounts (NLV < $ 25k). Again, errors could occur in some circumstances during live trading. On the other hand, this rule does not exist for European markets.
I am aware that QuantConnect cannot model all regulatory requirements from all countries, but it would at least be good to be able to see in advance what requirements an account must meet in order to license an Alpha for live trading without restrictions and without errors.
Are there any concrete plans to do something in this regard?
3. Lack of standardization in the Alpha Market
The criteria for approval of an Alpha have changed over time, but old Alphas have remained unchanged. This seems a bit unfair to the developers who joined later.
I’d appreciate it if all Alphas in the market have to meet the same minimum requirements.
Developers of older Alphas could be given some time to adjust their strategies if necessary so they can stay in the market.
But if they don’t meet the criteria after the deadline has passed, they should be archived. So you would still be able to see them for transparency reasons, but they would no longer be licensable/investable.
4. Minimum activity level of Alphas
Currently, at least 100 insights are required. However, insights do not have to result in orders. And also the specification of a minimum number of orders is not effective from my point of view, because then one could simply execute pseudo trades (buy a single share and then immediately sell it), so that the turnover remains low, but the number of orders increases quickly.
I think a better approach is to require a minimum portfolio turnover or a minimum average asset sales volume. That would be harder to manipulate.
Further suggestions:
- Minute resolution data or higher (provides more realistic order fills)
- No order fills at stale prices and Settings.StalePriceTimespan should be reduced to 1 minute (I think default is 10 minutes)
- Largest drawdown in backtest should be less than 30% (I am surprised that something similar is not already included)
- Recovery period for drawdowns > 10% must be less than 12 months (currently < 6 months is required, but I think this is too restrictive)
- PSR > 50% for each rolling 5-year period and SR > 1 for total period (I already commented on this on Slack. Using a rolling metric would also have the advantage of making the result less dependent on the choice of start date. In particular, the criterion “Backtest start date cannot be earlier than 7 years ago” should be omitted. A longer backtest is basically better because it covers more different market regimes.)
- Turnover > 20% (or an analogous criterion using average asset sales volume ratio)
- Skin in the game should be rewarded. Quants who are invested with a significant amount (e.g. > $ 10k) in their Alpha should be given special recognition in some form. Skin in the game is an easy way to spot good things.
- Manual adding of securities should be forbidden (no hardcoded tickers) to avoid selection bias. (Currently not feasible because for ETFs you need some background info like a list of constituents, asset class of constituents, weighting of holdings in ETF etc.)
- After approval the Alpha should maintain a good reconciliation score (e.g. DTW < 20%)
Let me know what you think.
Someone once mentioned that the overly restrictive requirements were one of the reasons why Quantopian failed, and I think the person is right about that.
To be more precise, we should very well set high standards, but not in terms of backtest performance, but in terms of resilience and closeness to reality.
Thank you for your time!
Vladimir
Arthur Asenheimer,
You are not the first to bring up the subject of a slight relaxation of performance criteria for the Alpha Market.
Mikko M posted his request some time ago.
I supported him, as I support some of your propositions:
- all Alphas in the market have to meet the same minimum requirements during life span
- Largest drawdown in backtest should be less than Compounding Annual Return
- but it would at least be good to be able to see in advance what requirements an account must meet in order to license an Alpha for live trading.
- The criteria for approval of an Alpha have changed over time, but old Alphas have remained unchanged. This seems a bit unfair to the developers who joined later.
But I strongly against your other suggestion:
- leverage = 2
it not allowed on retirement accounts (~ 70% retail traders).
- Most US-ETFs are not tradable for retail traders from Europe,
As I understand, you want to exclude them from Alpha market,
but what about 350 M possible retail traders from USA?
- Largest drawdown in backtest should be less than 30%,
- Recovery period for drawdowns > 10% must be less than 12 months
- PSR > 50%
The last 3 of your propositions are not relax a bit but request to open the door for garbage.
- Manual adding of securities should be forbidden.
This recommendation does not make it easy to create the most efficient rotation strategy
between different asset classes (stocks, bonds, commodities, real estate, currencies,
cryptocurrencies ...).
-> Someone once mentioned that the overly restrictive requirements were one of the reasons why
Quantopian failed, and I think the person is right about that.
I was a Quantopian hedge fund author and spend there more then 5 years.
Last two years I have access to quantopian-black.slack with 30+ authors from all over the world and ~10 from high level Quantopian's staff.
To my mind Quantopian failed on two reasons.
One similar to your proposition I last mention - requirement to trade only from QTradableStocksUS - universe of around 2000 stocks(no ETF or other securities).
The second probably came in 2017 from a major customer Point 72 - requirement to be dollar neutral (long-short), sector neutral, factor neutral...
Here is example of neutrality test on 19 factors and returns by each factor.
Here is example of specific and total Sharpe ratio test for each of the 14 days post-transaction,, specific and total return percentiles.
To my mind 60% for 6 years on portfolio of 600 stocks is not enough, but at Quantconnect you can do 4 times more with the same risk metrics just using 2 ETF's.
Quantconnect's current requirements are much softer.
Arthur Asenheimer
@Vladimir
I already know the discussion started by Mikko M and I agree with him.
But the discussion has unfortunately not achieved much.
However, I have now become aware again of Jared’s suggestion of a ‘frozen backtest’ approach. I think this is a good idea and would be a step in the right direction. But that alone is not enough.
You suggested there the inclusion of Information Ratio. But that would bring an additional problem, which is to determine an appropriate benchmark. For a strategy that trades US equities only, SPY might be a good choice, but what about a strategy that trades commodity futures?
What benchmark do you use there? SPY is no longer an appropriate choice then and using different benchmarks would make it difficult to compare Alphas.
>>"But I strongly against your other suggestion:
- leverage = 2
it not allowed on retirement accounts (~ 70% retail traders)."<<
Then what about short positions for stocks or FX trading? Both are restricted for IRAs afaik.
I’m not from the US so I’m not an expert in this topic, but the way I see it there are different type of IRAs and also differences depending on whether NLV is more or less than $25k.
As you can see, it quickly becomes confusing and complicated.
After all, you only allocate a portion of the available capital to an Alpha, so using leverage > 1 in the associated Alpha doesn’t necessarily lead to the need of trading on margin.
Additionally, it’s a requirement which is only of interest to US customers.
QuantConnect is already strongly focused on US markets and that’s fine. But I hoped there would be efforts to internationalize the business and that we would soon be able to trade non-US markets via QC Cloud.
My suggestion: instead of requiring all Alphas to be compliant with IRA regulations, it would be easier to leave it open and mark/flag an Alpha appropriately if necessary.
>>Most US-ETFs are not tradable for retail traders from Europe,
As I understand, you want to exclude them from Alpha market […] ? <<
No, definitely not. I didn't suggest anything like that.
I quote, "it would at least be good to be able to see in advance what requirements an account must meet in order to license an Alpha for live trading without restrictions and without errors."
>>The last 3 of your propositions are not relax a bit but request to open the door for garbage.<<
I don’t think you got my message.
The problem is that too much attention is paid to backtest/IS performance.
It wouldn’t, as you say, open the door for garbage, quite the contrary.
“There is a strongly negative linear relation between performance IS and OOS, indicating that the more we optimize IS, the worse will be the OS performance of the strategy."
(Freely quoted from the paper “Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance” by Baily, Prado et al., see screenshot at end of this comment.)
And yet IS performance is the key criterion for the Alpha Market.
And that aside, you should quote me correctly. I did not say “PSR > 50%”, but “PSR > 50% for each 5-year interval (rolling metric, incl. overlaps) and SR> 1”. The latter may even be harder to achieve than “PSR > 80% for the entire period”. I should have mentioned that I am also in favor of extending the minimum backtesting period to 10 years. Then you would have at least 7 intervals with 5 years each, which you could evaluate.
However, the most important advantage is that the results would then depend less on the choice of the start and end date.
>>Manual adding of securities should be forbidden.<<
This recommendation does not make it easy to create the most efficient rotation strategy between different asset classes (stocks, bonds, commodities, real estate, currencies, cryptocurrencies ...). "
As I said, this is currently not feasible anyway.
I quote, "Currently not feasible because for ETFs you need some background info like a list of constituents, asset class of constituents, weighting of holdings in ETF etc.".
However, the fact that adding securities manually always involves the risk of selection bias.
Arthur Asenheimer
Obviously, I meant : The fact is that adding securities manually, always involves the risk of selection bias.
Jared Broad
Thank you for the feedback 🙏🙏. Some direct responses:
> What benchmark do you use there?
We use 1.0 Sharpe Ratio to have a standardized multi-asset benchmark.
> Leverage
The multi-asset approach is precisely why alphas need to be 1.0 leverage, and we can apply the leverage in post-production - portfolio creation. This keeps it standardized across all alphas. Eventually, leverage would just be a GUI setting at deployment time; like in funds.
> Internationalization
It's not “desire” that slows this we just need to get it right in the US before considering international markets and fully supporting funds/investors. We're still not close to getting it right in the US. We need to focus early and diversify (flag etc) later.
> ETF Constituents:
Coming this week. Datasource is already in production, we're writing the docs for it now.
==========
Generally; I appreciate the feedback but broadly speaking our challenge today is not alpha creation it's raising assets under management/business and marketing challenges. No matter how awesome the alphas are without full allocations from investors they're underutilized. There's no shortage of investors in the US so the reason we're not at 100% capacity for all alphas is marketing, ease of use, trust in alphas, etc. Generally non-technical.
So we're investing in getting the word out there, building alpha-portfolio systems for start-up funds to harness alphas.
I agree with most of your points Artur is just worrying about the icing on a very large cake... I'll be happy to address once we've fully allocated the top 5 awesome alphas. If the average RIA pitching a Black Rock 60/40 ETF can manage billions we've got a few lessons in distribution to learn =)
The material on this website is provided for informational purposes only and does not constitute an offer to sell, a solicitation to buy, or a recommendation or endorsement for any security or strategy, nor does it constitute an offer to provide investment advisory services by QuantConnect. In addition, the material offers no opinion with respect to the suitability of any security or specific investment. QuantConnect makes no guarantees as to the accuracy or completeness of the views expressed in the website. The views are subject to change, and may have become unreliable for various reasons, including changes in market conditions or economic circumstances. All investments involve risk, including loss of principal. You should consult with an investment professional before making any investment decisions.
Vladimir
-> Obviously, I meant : The fact is that adding securities manually, always involves the risk of
selection bias.
In my opinion, if someone manually picks TSLA, GME, or other stocks after a huge take off,
it's definitely a selection bias.
But if you have chosen an index instrument like SPY, DIA, QQQ or others that is well managed
with a long history and your selection is based on statistically proven correlation between
market participants, this is not a selection bias, it is a choice of your strategy.
Vladimir
Jared Broad,
Good news!
Arthur Asenheimer
Thanks for your answer, Jared.
> We use 1.0 Sharpe Ratio to have a standardized multi-asset benchmark.
This applies to the calculation of PSR, but I was referring to the information ratio at that point. There the annualized performance of the benchmark (by default SPY) is used as you can see here.
> The multi-asset approach is precisely why alphas need to be 1.0 leverage, and we can apply the leverage in post-production - portfolio creation. This keeps it standardized across all alphas. Eventually, leverage would just be a GUI setting at deployment time; like in funds.
I assume this only applies to equities, right? If I create an algorithm with Futures and Alpha Streams Brokerage as brokerage model, then I am able to hold positions that exceeds the NLV. Technically we have here leverage of security = 1, but margin requirements are less than 100% so de factor leverage > 1 is in use.
I hope it stays that way, because changing the margin requirements for Futures makes little sense in my view, as they are set by the associated exchange and the use of leverage (along with liquidity) is one of the main reasons for the popularity of Futures.
> ETF Constituents:Coming this week. Datasource is already in production, we're writing the docs for it now.
This is great news! Looking forward to it. :-)
> There's no shortage of investors in the US so the reason we're not at 100% capacity for all alphas is marketing, ease of use, trust in alphas, etc.
I agree with you 100% on this one. The potential of the Alpha Market if far from exhausted. I think the three points mentioned (marketing, ease of use, trust in alphas) are indeed crucial, so I am very happy to see that you guys want to focus on that now. 👍
Regarding ‘trust in alphas’, the topic ‘risk of overfitted/biased strategies’ will still play a role, so I hope we will come back to this later.
Arthur Asenheimer
The material on this website is provided for informational purposes only and does not constitute an offer to sell, a solicitation to buy, or a recommendation or endorsement for any security or strategy, nor does it constitute an offer to provide investment advisory services by QuantConnect. In addition, the material offers no opinion with respect to the suitability of any security or specific investment. QuantConnect makes no guarantees as to the accuracy or completeness of the views expressed in the website. The views are subject to change, and may have become unreliable for various reasons, including changes in market conditions or economic circumstances. All investments involve risk, including loss of principal. You should consult with an investment professional before making any investment decisions.
To unlock posting to the community forums please complete at least 30% of Boot Camp.
You can continue your Boot Camp training progress from the terminal. We hope to see you in the community soon!