Hacker News new | ask | show | jobs
by avn2109 3564 days ago
>> "...impossible..."

https://en.wikipedia.org/wiki/Renaissance_Technologies

"...famed for one of the best records in investing history, returning more than 35 percent annualized over a 20-year span..."

4 comments

People have done it before. It has always turned out to be luck. Fantastic track record until they cease being lucky.

So, cynicism and economic orthodoxy aside, that sounds like a really cool company. Has anyone tried just tossing a big dumb neural network on stock data and investigated whether it can make money? It sounds very obvious, but a quick googling returns little. But I guess the investment industry is pretty secretive by nature.

> Has anyone tried just tossing a big dumb neural network on stock data and investigated whether it can make money?

... yes.

One of our Solutions Architects wrote this up for amusement: https://cloudplatform.googleblog.com/2016/03/TensorFlow-mach...
You've no idea what you're talking about.

Renaissance technology is a quantitative trading company, meaning they use computers to trade; they probably do thousands of trades per day, and consistently make a profit. There's approximately 0% chance that it's just luck.

Now, you might say, they were lucky to stumble upon a strategy that works. You can also say that the strategy will stop working at some point (because of competition), so their "luck" will run out, and they might not get "lucky" in time to find a new strategy. But their past performance was most definitely not just "luck".

Btw, in most other professions (e.g. the arts, technical inventions, sports) we call this kind of "luck", "skill".

RenTec makes too many bets for their track record to be just luck.
Are you seriously asking if hedge funds that pay hundreds of thousands of dollars salary to top experts from all kinds of fields have investigated using machine learning?
Obviously someone has, but I want to know the results.
You can't just "toss a NN to stock market data" and expect good results. There's too little historical data, you can't easily "generate" more data for the network to learn from, making it really easy to overfit. In other fields (e.g. computer vision) a lot of research has been focused on inventing techniques that prevent overfitting, thus enabling "learning" (i.e. generalization of patterns), such as dropout, convolutional neural networks, flipping/rotating images, etc. Very few of these techniques can be applied generally.
Dropout can certainly be applied generally - it's useful as a regularization technique (especially in wide and deep networks) to combat overfitting in other fields than computer vision.

CNNs can be used in other fields as well.

It turns out to be rather difficult, because in most setups you see in the stock market, the market goes up roughly half the time, and down roughly half the time.
There are millions of investors. The chances of a few outliers getting a long streak of returns are pretty good, even if they picked stocks at random. It's impossible to tell in advance who those lucky parties will be. But their existence is very likely.
Which is still less than someone who bought dell stock at IPO and held for 9 years.
So?

Dell's market cap at IPO was $80M.

You can't just pick one good company to invest in. What would these funds do with the other, oh, $10Bn they need to invest?

These funds are not getting that kind of returns on 10B. Often you see "fund not seeking additional investment."
That's my point. When you have so much money to invest you (are forced to) capture a more representative slice of the market and regress closer to the mean.