That all sounds right. “Fundamentals does get mentioned early but in complete isolation and dropped. I would not be sure that they used no fundamentals at all. News readers gets mentioned.
What was evident is that Frey used a lot of pair training and statistical arbitrage.
Part of the reason Frey’s model did not work was errors in the code. One error I remember was using a static number (no longer accurate) for the value of the SP500.
I just started the RIEF section.
The machine learning is interesting to me.
KERNEL regression is just nonlinear regression. There should be no debate about how that could be useful. And actually we know it was useful from the book. I can confirm personally it does just what the textbooks say it will do.
Has hidden Markov models been largely replaced by Deep Learning? I will defer on this.
The book should settle all issues as to whether statistics can be useful. I would be the last to force anyone to use it, however.
The SELF LEARNING …That blew me away. Anything I would say would be speculation.
If you ever get the ability to use Python within P123 you should try Kernel Regression.
This is one way to address the nonlinearity in the data. Read the book if you are not sure about this.
De Prado also sees great problems with methods that do not address nonlinearity. His view is that these methods are DOA.
How much it helps will simply depend on how nonlinear the data is without any need to reference RT or de Prado for support, however. This is just common sense.
If you cannot, somehow, connect your data to Python then I cannot think of anything usable from the book.
Nothing usable other than knowing retail investors (clearly not pension funds as per the book) are their revenue source. We compete with RT. We compete for profit opportunities and sometimes we are their target.
The Medallion Fund was named because of the prestigious awards the mathematics Ph.D.s had received. In contrast some at P123 literally state the math cannot be useful. Somehow they know this and may never have bothered to calculate a p-value. The p-value guides every decision regarding methods at RT, according to the book.
Maybe RT will be proven wrong ultimately—although their results are not a strong argument against their methods. I am just saying their ideas do not transfer well to P123 at this time.
BTW, it is not just Renaissance Technologies. D. E. Shaw & Co does the same things. The same D.E. Shaw & Co. for which Larry Summers served on the board. The Harvard trained Larry Summers who served in the Clinton administration (US Secretary of the Treasurey), the Obama administration and the US World Bank. Whether these methods can work (or not) is not controversial–except at P123 it seems.
I tried it back in '16 and decided it wasn’t for me as it seemed geared towards day-trading. But I didn’t use their AI assist feature, just the screener. However, one investment Meetup member thought it was great and claimed to be making money from it. YMMV.
Yes thanks Walter. mmssand, my apologies for missing you point. I did not know Trade-Ideas was a Web site. Interesting and I look forward to responses. Sharing the thread with an RT concern….
Does anyone think that what we are noticing could be caused by RT, D.E. Shaw & Co. and others.
I know RT said their machines are seeking-out Dentists. While Eye Doctors were never mentioned as targets my paranoia has me wondering.
In other words, have the machines noticed what—we at P123—do in aggregate. Are we big enough for the machines to notice. If not, are we, Zachs, Quantopian and others big enough to be noticed by the machines and be targeted?
Not evidence, but the cyclical nature of this seems to have gotten better just at the time JudgeTrade, Yuval, I ,and others I think, have changed their algorithms. Or said so in posts (I just said so).
I have changed too: cutting the percentage and total amount invested in my ports mainly. So not a comment on how people should handle unexpected behavior in their ports. But just wondering if the machines might be adapting and moving to easier prey now—general surgeons are slow and are easy prey. There is a lot of money controlled by these firms: enough to be responsible for some of the cyclical nature of the market?
Just my paranoia perhaps and I have not even seen that new Terminator move yet!!!
Thanks yuval. That was my take away … I dont play their game, and I dont want to try. I have also heard it speculated that the real Ren Tech advantage is they were the first on the market of their kind so they came along early and got exclusive rights to proprietary datasets locked up on 100 year contracts. If there was just some sort of math or ML special sauce someone would have emmulated it by now.
There was a brief time in the early '90’s where technology, computing power, was making a breakout. You had to be aware to see it. In my case I built boxes from scratch to stay ahead of the curve, Back then you could get an easy edge by accessing data downloads few others could get to.
I had a satellite dish which delivered tick data. I ran neural networks using software that was delivered on floppy drives. For several years there was a window where if you could stay on top of the technology breakout you had an edge,
RT was there at that time. If you read the book carefully you will learn that the seed capital came from a lucky bet in the futures market…
The book says they have achieved a 66% gross return, and somewhere else I saw they usually have 4-5x leverage. So does that mean they have an unlevered return of 16.5% and then levered up 4x to get 66%?
Maybe someone has some actual data and experience on what corners of the market RT is not playing in.
I suspect, but do not know, some people playing in the micro, micro-caps (even OTC for some P123 members) succeed because RT, D.E. Shaw & Co, and other do not play there. So I am not debating that such places may exist—in fact I suspect they do.
But the only place you can be sure that you are safe is ETFs and long term investments. And long term I will not beat Marc and the Professional Analyst. A man has to know his limitations.
If you do decide to play in their playground you better have an advanced version of some of their tools. Even then you would be picking off some retail investors along with them. You would not be winning against them. Kind of like playing poker and you are not usually the big winner each night but you are not the guy who loses every night either. Both you and RT would be working on that losers each poker night.
I would like to play with some of their tools. But more than likely I will be (mostly) moving to ETFs.
BTW, if you read the book you will find RT does have some funds (not the Medallion fund) designed for pension funds that do long term investments and these use fundamental data. The book does not say that the Medallion Fund uses no fundamental data. Analyst estimates is mentioned as a useful signal but whether the fund uses that now is not known to me. It is safe to say fundamentals are not the biggest part of the Medallion Fund algorithm, I think.
Their average holding period is 2 days. No doubt they do some day trading but if the average is 2 days that puts some of their trades into our space. Meaning same trades may have holding periods similar to some of our weekly rebalance ports.
If it works they will use it. If an automated method works for you it will work for them. I am pretty sure that if you find a fundamental signal that works for you they will be using it in one of their funds.
Yeah, what their actual unlevered annual return was has always been a question of mine too. I know employees were also able to convert their Medallion Funds into employer provided Roth IRAs during a loophole in 2012, so the returns are tax free too.
Levered is important. But the reason the can lever so much is that they use pairs trading and statistical arbitrage. Meaning about half of their portfolio is short. They do not necessary have a positive expected return from the short positions.
If they went long only…who knows?
Once you start doing pairs trading, comparisons to what we do gets difficult if it has any mean anything at all.
Ok so assume for a second the short positions were not for alpha, but just as a hedge.
Once you find your long position to generate alpha, how do you select the corresponding stock to short? You could use past correlation but is that at all persistent?
I know Thorpe was working in this and he found that that stocks in the same sector should be used. One could speculate that the decide somehow to generally mix coke and Pepsi or two drilling stocks.
The book said they were looking for stocks that had a big buyer (or seller) that pushed the price up (down). Point being it is not really a bet on soft drink sales, say. So they might buy coke and sell Pepsi with no opinion on the fundamentals.
Probably too much speculation in that already. Wish I knew more.
Marc calls an image from the book: “brief and flawed anecdotal table.”
This was for a table with the NET gains of the Medallion Fund reported in the book. I won’t reproduce it again here.
BTW, this is what I found about the RIEF fund on the internet. I missed it in the book (or it was not there). I will paste the entirety:
“Its Renaissance Institutional Equities Fund (RIEF) LLC Series B also had envying returns in recent years. In 2013 it delivered 17.64%, in 2014 it brought back 14.53%, followed by 17.37% in 2015, 21.46% in 2016, 15.22% in 2017, and 8.5% in 2018. The fund’s total return amounted to 319.31%, for a compound annual return of 11.5%, while its worst drawdown was 34.58%”
The fund was created in 2005 according to Wikipedia (I was interested in when the drawdown might have occurred).
I believe the fund may use shorts and/or use calls/puts. But there was recent article about Ray Dalio doing this with his Bridgewater Associates. So the success of quantitative strategies needs to be taken seriously with regard to longer term investing, I think.
Also Ray Dalio is free to look at any alternative data he wishes to and I suspect he might. Even a fundamental analyst might want to look at the data FactSet provides regarding Credit Card data, for example.
My guess would be that, since they are both designed for pension funds Bridgewater Associated and RT RIEF would play by the same rules. It would be interesting to make a direct comparison to Ray Dalio’s fund. I have not done this.
I hope to have nice returns like that using a single ETF model I developed here recently, influenced in part by Georg’s models. Thank you, Georg! In backtesting it outperforms the Renaissance results 2013 through 2018. It holds a single unlevered large ETF chosen from a small set using a ranking system. It doesn’t use hedging. And it doesn’t use market timing except within the ranking system. At this time it is only a model with no OOS results, so time will tell. I am now using the model with real money.
I had an extensive career in IT and have coded in machine language, assembler, basic, DOS, Fortran, PL/I (my favorite), a bit of COBOL, C, and C++ (and maybe others I don’t remember). I don’t care to ever code that way again unless forced. In addition, machine learning can be helpful to uncover hidden correlations but it is incapable of assigning real cause and effect relationships which are very important to determining likely OOS impacts. I struggle to maintain a proper view without using AI and it could be impossible if I let myself be enthralled “by the dark force”!
I have personally witnessed extremely poor real business decisions made by others as a result of assuming an artificial intelligence model accurately forecast the near future without taking into account known environment changes that would make the model absolutely untenable. AI can do that very easily. I am probably going to miss out on some excellent AI approaches, but I prefer to stick with an approach that parallels the scientific method. Observe, hypothesize, experiment, and prove or disprove, then rinse and repeat. I don’t care to emulate what RT does. Except what I am modeling may be somewhat similar to their approach in some ways. Using ETFs prevents the easy use of fundamentals, and from what has been mentioned in this thread I suspect RT operates primarily with recent technical data.
Why wouldn’t you optimize an ETF model up until 2015 then use 2015 till now to get out-of-sample results. That way you would not have to wait several years to know if the model is good.
If the model needs to be “trained” in some sense you can always retrain it with the latest data before you put money into it. For a port you could optimize again on the latest data.
StaveA re-optimizes.
This is what the AI expert P123 hired was paid to do.
Some have very correctly said they cannot do that with stock models because they already know what has worked from 2015 until now from their other models and other backtests. I think this as called data snooping. They are 100% correct about this so it cannot always be used.
Some signals, especially technical indicators do not have this data snooping problem—depends on the person. I for example would not know whether a 30 day moving average would be better than a 45 day for the S&P 500. For Georg there may be data snooping with moving averages as he has a lot more experience with this than I do.
Just an idea that can be used for some (not all) situations, I think. Not too radical when it can be used because, as I say, P123 paid for this.