Art project: Train rats to trade markets | Hacker News Reader

Art project: Train rats to trade markets

artmarcovici.com

138 pointswaterlesscloud44 comments

The obvious next step is to try the same with voles, shrews, etc., and start a hedge fund that is actually located in a hedge.

^ this :)

By the way, there's a little "up" triangle next to each post which means "this" ;).

I'd suggest Hedgehogs :)

the bottom of the link says just that -- the best traders breed even better traders, so a hedge fund is inevitable

/edit

Dude has had an interesting life. Bio from his webpage:

Michaels career, not only as an artist, is a rather unusual one: born in 1969 in Vienna, he got interested in technology, especially electronics and mathematics, already when 7 years old. At the age of 12 he worked as a programmer. He quit school at 17 and started his first own business in the financial field, publishing analysis on the financial markets and managing funds, until he sold his business at the age of 23. The following 3 years he spent climbing and hiking in Africa, USA, Asia, and all over Europe. From 1995 on he was the publisher of werk-zeug, a technology and art magazine, as well as streetfashion, a magazine featuring fashionable people on the streets all over the world. He was also active in the field of software development, is the originator of many inventions, and holds international patents ranging from climbing equipment and bicycle gears to trading systems and electronic payment systems. Already in 2001, Michael decided to sell his companies to start his career in arts and works as he calls it. In the very same year though, he started a side business on eBay that eventually became the world's largest powerselling enterprise on eBay, a company with 80 employees and a turnover of 30 million euros a year. In early 2005, the company went bankrupt, and Michael lost all his money in the process. Never shying away from new challenges, he decided to write a book about the company and its end: English and German versions of "The end of EBay" are available through Amazon here. Qentis' bankruptcy also had its good sides: finally, Michael could concentrate on his art and private studies. Michael’s works are the result of a rich live and a long history of ideas. The artist about his work: "there is no way to (mate)realize all the plans and projects I have on my mind in one lifetime- I can only pick the best and feasible". His work touches a broad variety of fields such as technology, politics, science, social topics, and style.

Cute project, but since it isn't actually hooked up to live trading, I'm going to suspect that they're not actually dealing with the issues which would keep them from making a profit on those trades: trading fees, losses going into the bid-ask spread, the fact that large enough purchases actually move the market, and the fact that the crash which wipes you out can come along on years later just when the victories of the past few years made you think you had everything figured out.

Simulated trading is all well and good for the long-term buy-and-hold sort of trading where you can assume you're a medium-small-sized fish in a huge pond and these errors are inconsequential, but with ultra-short-term trading like this the texture of the market matters a lot more.

That said, if you're a day trader of some sort trying to compete with the task the rat has been assigned to do, there's a very much elevated chance that you're in the wrong business.

I seem to recall that the most likely predictor of whether the next tick in a liquid market will be up or down, is whether the last tick was up or down.

In other words, if the last tick was upwards, there is a >50% chance that the next tick will be upwards too.

A quick'n'dirty Google search turned up some research: http://www.researchgate.net/publication/234834258_Random_Wal...

Looking at the ticker tracks the artist has used, I wouldn't be surprised if the rats are effectively being trained to Buy when they hear high-pitched notes, and Sell when they hear low-pitched ones.

It's actually the other way around. For most stocks, if the last tick was upwards, the next tick is more likely to be downwards (where a "tick" is any time the best bid or ask moves, and "price movement" means the change in the mid price).

For example, taking a random stock over a random three month time period, I compute that after an uptick, the probabilities for the next tick are -

  P(uptick)    = 39.3%
  P(downtick)  = 56.3%
  P(no change) =  4.4%

where "no change" typically happens when the best bid and ask change by the same magnitude, but in opposite directions.

Of course, this is not very useful for trading, because it doesn't give you enough of an edge to overcome the bid-ask spread.

Okay, here's the breakdown of the five sample ticker tracks provided by the artist.

                                                    NR231   NR287   NR320   NR440   NR442   Totals
  Uptick followed by an uptick (UU)                    6       5       9      14       5      39
  Uptick folowed by a downtick (UD)                    2       4       3       3       4      16
  Uptick followed by no significant change (UN)        1       0       0       2       0       3
  No significant change followed by an uptick (NU)     1       1       0       2       1       5
  No significant change followed by a downtick (ND)    2       1       0       0       1       4
  Downtick followed by no significant change (DN)      2       2       0       1       2       7
  Downtick followed by an uptick (DU)                  1       4       3       0       4      12
  Downtick followed by a downtick (DD)                 8       6       8       1       6      29

There were 58 upticks in total, 39 (67.24%) of which were followed by another uptick.

There were 48 downticks in total, 29 (60.42%) of which were followed by another downtick.

There were around 480,000 ticks in my sample - who do you believe? ;)

The difference is probably in what is defined as a "tick" for the purposes of computing the stats. If you are using traded prices as opposed to quotes, for example, you will get very different results. If you sample at regular intervals (e.g. 1s) you will get different results again.

A lot of people didn't get the idea that this is an art project, not an in-earnest investment strategy.

Yeah, this isn't real, is it?

It's an art project, not a science experiment.

Was the training and price input necessary at all?

Doesn't random generator outperform some hedgefunds?

One of the rats achieved a 53.5% success rate (the author did not specify whether this was over the entire 800 data points or only the last 100).

When tested on 800 data points, a purely random strategy would have a > 53.5% success rate about 2% of the time.

Finally. We can replace the Wall Street folks at 1 / 1.000.000 the price of today. This will be a significant boost to the economy.

LOL, you under estimate the cost of the cheese... and the dropping removal service... but yes, for the most part, I think you are right!

Imagine the savings on cocaine alone!

no no, the fine print also show the rats were "conditioned" with a little bit of magic powder...

Someone said hedge fund managers don't run on coke and models any more. But hey, now it's coke the beverage, and mathematical models...

I am no AI expert, but couldn't a modern AI be trained more or less in the same way?

"AI" is not a single thing. There are some projects that you train to respond to signals, sure. Some of them are already in use in HFT. Modern trading tends to use a lot more inputs than just the historic price though - you also look at wire services, twitter, etc. So I wouldn't expect this to be competitive without having access to more data.

Wasn't the point that the rats only reacted the the live feed, and didn't have any ancillary data? That ticker tracks converted to sound were enough to perform strikingly well in terms of deciding holding moves?

...the training was almost finished; the performances of the top 4 rats had turned out to be comparable to those of the world' s best fund managers. Their ability to recognize sound patterns generated from the market's ticker tape was incredible. And the rat traders also hold another advantage: unlike their human counterparts, they are not likely to be distracted from news or economy fundamentals, their own personal or their bank's financial status.

A random number generator outperforms most fund managers...

Indeed, a chimpanzee was actually the top performer in 1999, by throwing darts at ticker pages. Bad year for the market.

http://www.marketwatch.com/story/paul-farrells-commentary-ch...

http://www.dailyfinance.com/2011/08/03/in-honor-of-planet-of...

I should point out that it's not at all surprising that the "top 4" rats performed very well. If you got a couple of hundred people, animals, or random number generators to trade using any set of rules you like, and then looked at the top 4, you would almost always find that they outperform, because that's what being in the top 4 means!

The real question is whether the same four rats would be the outperformers if you ran the experiment again.

I guess all we can do is hope that we can become Mr Kleinworth Morgan Jrs. =D

this is a mix of neural network and genetic algorithm.

the rat animal for the right job...

next: hedgehog fund managers

If rats can actually beat human traders, why shouldn't we put the traders in cages and let the rats have our money?

this is fascinating. i love the mix of scientific experimentation with an art piece. i'm wondering at which point the improvements in the newer generations will taper off... maybe a rat could be the next warren buffet?

Rats? Is there a pun intended?

Rats Awesome Trading System?

The fat cats aren't going to like this...

Rats Are The Solution

Seeing it's an art project, I see the "rat race" analogy.

When the project tanks, everyone will be the first to leave the sinking ship.

Perhaps this explains the ebb and flow of the market?

Consider for a (rat) trader:

- success

- get treat

- eat treat

- loop

... passage of time...

- get fat

- judgement clouded by excess fat (perhaps inhibiting movement)

- less successful

- less treats

- loses weight

- become more successful again

...and so on...

Silicon valley "rat trading" funding announcement in 3... 2... 1...

this is so interesting, one more reason to have a rat as my pet.

Is this a joke or not?

http://www.5markets.com/

+100 LNG RATS

The sonorification bit has me interested. Perhaps a human trader could outperform the rats, or at least themselves, if they went just by sound data without trying to think too hard? I can see he's going to try training humans to trade this way, the results of this will be pretty interesting.

It may very well turn out that the difference between trading by looking at numbers and trading by sound and intuition is like looking at someone throw a ball and predicting where it will land versus being told the ball's mass and initial vector - in the former case you can very quickly predict with a good accuracy where it will land, while in the latter you need a good amount of time and a few sheets of paper and you achieve accuracy that's not that much better.

And yes, I realise that once you stretch the analogy too far it doesn't work, the ball will follow a completely deterministic trajectory, while the stocks move in a rather random fashion and the principles (if you can call them that) that govern their movement are not known to nearly the same precision as the principles of motion of a rigid body on Earth in normal atmospheric conditions.

But I'd say that if you look past that, the analogy is informative and serves to roughly illustrate the point.

Reply on news.ycombinator.com