Are Renaissance Technologies just trend followers?

[opihiman2]

The new book is great, and it's illuminating.  Sad to see how much errors there are in this thread about RenTech. 

1) Medallion hasn't been publicly available for a LONG TIME.  Their institutional funds are and have very public performance data.  However, the institutional fund performance #'s are nothing to write home about.

2) Medallion is limited to around $10B, I think.  They disburse all gains above and beyond to stakeholders.  Even at this size, their performance #'s are unreal.

3) Buffett doesn't even come close to Medallion #'s.  He wishes he could.  Even if he were trading with a small AUM, I HIGHLY doubt young Buffett would come close to Medallion Fund #'s

4) Medallion isn't really stat arb.  Not even close.  They did do some stat arb before, but not with equities.  They were pretty good in commodities and debt investments.  But, what really ramped up their #'s since late 90's was pattern recognition software.  They hired two guys from IBM from their speech recognition division who finally was able to get Medallion a significant edge in equities.  As a side note, one of the guys is a far right conservative who likely helped Trump win the election. 

5) They use leverage but not LTCM levels of leverage.  They can scale in and out of leverage fairly easily. 

6) Jim Simons gets way too much credit for the success of Medallion.  His main contribution was probably the idea of using quantitative models for trading and team gathering. 

7) They are definitely not fundamental investor types.  At least not since the early 90's.

[RuleNumberOne]

Edit: the links don't work. Need to Google for author-name followed by "dblp"

Robert Mercer's papers are here: https://dblp.uni-trier.de/pers/hd/m/Mercer:Robert_L=

David Magerman's papers are here: https://dblp.org/pers/hd/m/Magerman:David_M=

No doubt these ideas were used at Medallion

"Jelinek wrote, "The performance of the Renaissance fund is legendary, but I have no idea whether any methods we pioneered at IBM have ever been used. My former colleagues will not tell me: theirs is a very hush-hush operation!"

If they didn't use any of it, they would have told their lead co-author Jelinek.

[RuleNumberOne]

I know very little about neural networks, but the speech recognition research that happened in IBM in the 1990s was antiquated long ago. But suppose we start with what kind of queries can neural networks answer that would be relevant to stock trading?

Fill-in-the-blanks.

Given a sequence of winners and losers for training the model, the neural network could predict missing winners or losers. E.g. given the set of today's winners, which stocks are missing from the list... 5000 stock tickers per day over 10 years would be a total of 12.5 million words for training.

They would still need to add fundamental data as input to justify why a stock flips from reliable winner to loser and vice versa.

If multiple stocks are missing from today's stock market winners, there are likely no outliers, as opposed to a single stock missing from the list of winners...

[RuleNumberOne]

I think they also use some sentiment analysis of analyst reports or news articles for drug companies. Drug approvals in the case of NVO and VRTX, merger in the case of BMY and CELG.

The appearance of CMG could be explained by detecting an outlier in the sequence of winners.

I think the publication of this book will result in a lot of competitors. They have shown it works and there are a lot of smart people out there who can replicate these results - who never tried before because they had no idea such techniques would work.

[cameronfen]

Edit: the links don't work. Need to Google for author-name followed by "dblp"

Robert Mercer's papers are here: https://dblp.uni-trier.de/pers/hd/m/Mercer:Robert_L=

David Magerman's papers are here: https://dblp.org/pers/hd/m/Magerman:David_M=

No doubt these ideas were used at Medallion

"Jelinek wrote, "The performance of the Renaissance fund is legendary, but I have no idea whether any methods we pioneered at IBM have ever been used. My former colleagues will not tell me: theirs is a very hush-hush operation!"

If they didn't use any of it, they would have told their lead co-author Jelinek.

His research at least as of 2019 is predominately neural network based.  Transformers, LSTMs with attention that like what is hot now and what was hot 3 years ago in language models with neural networks.  He has some other stuff but I am sure his work is mainly Neural Network based.  I bet by now there is no way Rentech are not using transformers to forecast these time series based on his expertise and what works.  It is likely he ran a giant transformer with like maybe 1billion+ parameters on the time series of every single financial asset and with maybe slight modification is running the same transformer to predict movement in those given assets.  If I had to guess you can copy the structure of the largest transformer, megaton-ln, from here: https://blog.exxactcorp.com/megatron-lm-unleashed-nvidias-transformer-megatraon-lm-is-the-nlp-model-ever-trained/ , steal the transformer base architecture from here: https://github.com/tensorflow/tensor2tensor and with little coding knowledge but 2 or 3 million dollars to spend on AWS you can likely replicate 50% of there returns just by training it on every possible asset class time series.  My guess is that’s the core of the model and all the ml smarts in the world gets only somewhat marginal improvements from there. 

[RuleNumberOne]

Robert L. Mercer (the Renaissance guy) last published a research paper in 1995.

Robert E. Mercer is a professor somewhere and he is still publishing research papers.

Yeah, it is very easy now to run the latest algorithms on any amount of data compared to the 1990s. Lot of machine learning packages available.

But the key is what inputs to feed - how do you structure the inputs.

Edit: the links don't work. Need to Google for author-name followed by "dblp"

Robert Mercer's papers are here: https://dblp.uni-trier.de/pers/hd/m/Mercer:Robert_L=

David Magerman's papers are here: https://dblp.org/pers/hd/m/Magerman:David_M=

No doubt these ideas were used at Medallion

"Jelinek wrote, "The performance of the Renaissance fund is legendary, but I have no idea whether any methods we pioneered at IBM have ever been used. My former colleagues will not tell me: theirs is a very hush-hush operation!"

If they didn't use any of it, they would have told their lead co-author Jelinek.

His research at least as of 2019 is predominately neural network based.  Transformers, LSTMs with attention that like what is hot now and what was hot 3 years ago in language models with neural networks.  He has some other stuff but I am sure his work is mainly Neural Network based.  I bet by now there is no way Rentech are not using transformers to forecast these time series based on his expertise and what works.  It is likely he ran a giant transformer with like maybe 1billion+ parameters on the time series of every single financial asset and with maybe slight modification is running the same transformer to predict movement in those given assets.  If I had to guess you can copy the structure of the largest transformer, megaton-ln, from here: https://blog.exxactcorp.com/megatron-lm-unleashed-nvidias-transformer-megatraon-lm-is-the-nlp-model-ever-trained/ , steal the transformer base architecture from here: https://github.com/tensorflow/tensor2tensor and with little coding knowledge but 2 or 3 million dollars to spend on AWS you can likely replicate 50% of there returns just by training it on every possible asset class time series.  My guess is that’s the core of the model and all the ml smarts in the world gets only somewhat marginal improvements from there.

[cameronfen]

Robert L. Mercer (the Renaissance guy) last published a research paper in 1995.

Robert E. Mercer is a professor somewhere and he is still publishing research papers.

Yeah, it is very easy now to run the latest algorithms on any amount of data compared to the 1990s. Lot of machine learning packages available.

But the key is what inputs to feed - how do you structure the inputs.

Edit: the links don't work. Need to Google for author-name followed by "dblp"

Robert Mercer's papers are here: https://dblp.uni-trier.de/pers/hd/m/Mercer:Robert_L=

David Magerman's papers are here: https://dblp.org/pers/hd/m/Magerman:David_M=

No doubt these ideas were used at Medallion

"Jelinek wrote, "The performance of the Renaissance fund is legendary, but I have no idea whether any methods we pioneered at IBM have ever been used. My former colleagues will not tell me: theirs is a very hush-hush operation!"

If they didn't use any of it, they would have told their lead co-author Jelinek.

His research at least as of 2019 is predominately neural network based.  Transformers, LSTMs with attention that like what is hot now and what was hot 3 years ago in language models with neural networks.  He has some other stuff but I am sure his work is mainly Neural Network based.  I bet by now there is no way Rentech are not using transformers to forecast these time series based on his expertise and what works.  It is likely he ran a giant transformer with like maybe 1billion+ parameters on the time series of every single financial asset and with maybe slight modification is running the same transformer to predict movement in those given assets.  If I had to guess you can copy the structure of the largest transformer, megaton-ln, from here: https://blog.exxactcorp.com/megatron-lm-unleashed-nvidias-transformer-megatraon-lm-is-the-nlp-model-ever-trained/ , steal the transformer base architecture from here: https://github.com/tensorflow/tensor2tensor and with little coding knowledge but 2 or 3 million dollars to spend on AWS you can likely replicate 50% of there returns just by training it on every possible asset class time series.  My guess is that’s the core of the model and all the ml smarts in the world gets only somewhat marginal improvements from there.

Oh interesting.  Strange both are in NLP.  Also maybe should have checked because I was surprised a 70 year old is both recently a CEO as well as publishing papers as well as running a Koch style political funding arm.  Either way before I think 90% prob base model is Transformer.  Now probably 80%.  None of his stuff has anything to do with neural networks from a cursory look, which is unsurprising as ANN only got hot 2015.  Still, I’d be surprised if anything could outperform a core transformer model (megatron-ln) trained on basically every asset time series. 

[RuleNumberOne]

Until I read the book I thought quant traders didn't make money. It was an eye-opener. But it seems there have been a lot of people copying Renaissance over the last few years. Shouldn't the advantages that quants have get competed away?

WorldQuant manages money for or is owned by Millennium management (the same firm that two Renaissance people defected to as described in the book).

Citadel and Two Sigma are hiring computer science grads from universities and creating a campus environment.

https://www.bloomberg.com/news/articles/2017-03-06/citadel-joins-two-sigma-chasing-quants-in-campus-recruiting-push

"Two Sigma will take over “The Bridge,” a space on the new Cornell Tech campus on Roosevelt Island in New York, where engineers and entrepreneurs will work.

The hedge fund staff will collaborate with Cornell students and professors on machine learning and data science projects, creating a pipeline of academic talent to the firm. Job candidates will put on virtual-reality glasses to watch a video that explains how the hedge fund sees the world awash in data.

The students -- handpicked from 400 applicants -- are competing in a datathon hosted by the $26 billion hedge fund Citadel. Ken Griffin’s firm is upping the ante in the industry’s chase for data scientists and engineers, hosting 18 competitions at universities across the U.S., Britain and Ireland this year. The prize in the final data championship: $100,000.

Igor Tulchinsky, the founder of WorldQuant, is pitching a perk that breaks the tradition of hedge-fund secrecy. The $5 billion firm is hiring at least 15 teams of quant managers for its Accelerator platform, offering them the right to keep their intellectual property. The quant hedge fund has also opened more than 20 offices in 15 countries, including emerging markets like Russia and Romania, to find engineers. Talent is distributed around the globe, Tulchinsky said at the Milken conference, “but opportunity is not.”"

[fareastwarriors]

Renaissance hedge fund reportedly having one of its best years ever


https://www.cnbc.com/2020/04/17/renaissance-hedge-fund-reportedly-having-one-of-its-best-years-ever.html

[bennycx]

Looking at their pnl profile, they should be doing market making (agency trading), similar to investment banks but hidden in a hedge fund without all the compliance and operational costs.