"Since I did not want to make it too complicated, I only used the USD/EUR future to make the rats experts in this specific market segment, but other rats can be trained in other markets as well. I trained the rats for about 3 months, starting with 80 Sprague Dawley laboratory rats, 40 males and 40 females with the intention to cross the best of them to genetically create the best traders through select breeding."
It's clearly not a real experiment. The social commentary angle is interesting, although personally I'd tend to lean more toward elaborate hoax. The only thing I don't understand is why it has so many votes. It's not like it's at all insightful or illuminating when viewed as a thought experiment, so people must just be falling for it.
But on the other hand the author seems to be accomplished in the financial field. From his site:
"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.
[...]
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."
Essentially, he made a more complicated version of a moving average quant system.
The more interesting part was breeding top-performing rats to see if they are able to improve their "trading" genetically. You can do this exact same thing using evolutionary algorithms with trading systems that help you to adapt to the market.
Quant finance is fun when you're doing it for yourself, not a bank.
Obviously a positive return cannot be made from moving averages. But how do you know that the rats don't develop some sort of intuition that can make a positive return?
First, moving averages do play a role in some (profitable) quant systems.
Second, I never said that the rats couldn't create a positive return, but you could model the rats behavior quantitatively and develop a system, but that's just bringing it back full circle.
Well Keith Chen @ yale showed that capuchin monkeys show the same fundamental behavior as humans. By the end, they were trading jello beans for sex and creating their own bubble economy. http://tinyurl.com/au5hq
Comment from a buddy of mine Kartik, that I figured I'd share with the thread.
Fascinating. I'm extremely skeptical, of course. The thing that bothers me most is that they list percentages (perhaps this is the percentage of time that the rats guess correctly?) rather than $ amounts.
Even if the experiment fails, explaining the fact that the later generations performed better could lead to some really interesting science.
Percentages rather than dollars are worthless. There are skewed bets; you can predict "up" rightly 70% of the time and lose money on the bigger downward 30% of the movements.
Your response indicates that you did not understand the parent post.
Assume I have some system where I guess right only 25% of the time, but if I guess right I get $5. If I guess wrong I lose $1. My expected return is (0.25 * 5 - 0.75 * 1) = 0.5, which is 50 cents.
So, I'm doing worse than a coin toss, but still making money.
This is what the parent poster meant by a "skewed bet". The payoff for guessing "heads" or "tails" is not the same.
You can't just look at the results of the coin toss (correct guess, incorrect guess). You also have to look at how much you get paid for a correct guess, and how much you lose for an incorrect guess.
In most real-life situations, the payoff is not symmetric (equal win and loss amounts). This is why your "if you can't beat a coin-toss" comment is meaningless (and usually incorrect).
A. I guess correctly 25% of the time (via some method) and make money because of the skewed payoff. (0.25 * 5 - 0.75 * 1) = 0.5
B. I flip a coin and guess correctly 50% of time and make EVEN MORE money because of the skewed payoff. (0.5 * 5 - 0.5 * 1) = 2
Why should I ever go with option A?
Furthermore if your method lets you guess correctly 25% of the time, why don't you simply make the opposite trade and now you are guessing correctly 75% of the time!
Are we are talking about something fundamentally non-binomial? (buy, sell and do nothing or something even more complicated?)
Ah, ok. If there's a skewed payoff for B, then yes, you're correct. I read your post to mean that you were not considering the payoff, and only the probabilities.
As for the rest, what does "opposite trade" mean? For instance, going short versus going long carries very different risks. It's unlikely that your payoff would simply be mirrored.
> I read your post to mean that you were not considering
> the payoff, and only the probabilities.
Ultimately, the payoff is just acting as a constant offset to the break-even point (assuming the payoff doesn't vary with some other parameter). The skewed payoff may mean that you only need 25% accuracy to break even or it could mean that you need 95% accuracy to break even. It doesn't matter. Either way, you can effectively ignore it and consider, for a given payoff schedule, how your prediction algorithm will perform.
If we live in a universe where you can under-perform a coin toss and still make money because of how the bet is skewed, then I can do better by flipping a coin!
If the bet is skewed the other way, then we will both lose money but my coin toss will lose less.
Let me restate that: you can completely dissociate your prediction algorithm from your cost function
> As for the rest, what does "opposite trade" mean? For
> instance, going short versus going long carries very
> different risks. It's unlikely that your payoff would
> simply be mirrored.
Here my (lack of) knowledge of the various types of financial transactions that can be made puts me at a disadvantage, but the way the "article" describes it, the rats were trained to press a green button (long, betting prices were going to go up) or a red button (short, betting prices were going to go down).
There are only four outcomes here (as I understand it):
* Predict Up, Moves Up
* Predict Down, Moves down
* Predict Up, Moves Down
* Predict Down, Moves Up
It's possible that you can better predict upward movements than downward movements, but lets assume for simplicity that you (or the rats) are equally bad at both.
If you were able to predict at 25% accuracy, I would take what you told me (up or down) and flip it-- because you are actually performing at 75% accuracy, you just don't know it. Then I would make a trade. I don't know what the most clever trade that could be made based on that knowledge, or what the various payoffs associated with them are, but as I showed above, it doesn't matter for the analysis.
As far as I could see it was only Mr. Kleinworth Morgan Jr 5 which performed noticeably better. Run through a statistical test I doubt the results observed would deviate from random variation.
Issues with this that I see (aside from the obvious lack of statistical rigor) are that he gives the rats repeated input. Effectively, he is selecting for the rats who best memorize tracks 1-600.
I wonder how this compares to using a neural network. I see no comparison to using other methods to predict the market, though he claims to outperform humans.
The study is very eye opening to say the least. Results are based off of and emotion-free instinctive approach to choosing outcomes based on (to a degree) fundamental investment principles. Pavlov's application to investing. I agree with the thread, the most fascinating aspect is the generational improvement throughout the experiment.
I would say the % means that out of 100 investment decisions, the rats have been correct at x-percent. I have seen this before and it looks like work in progress, but maybe more art then science, the man is an artist...
The market makers, those tend to be human beings who buy and sell stocks, watch the market and continuously adjust their actions based on their observations of the market.
No elaborate theory here, just stating the oversimplified obvious.
A neural network, whether software-based, hardware-based, or wet-ware based (rat in this case, or human for that matter), can only spot patterns in a given set of data which are there. No patterns in the data means no predictability. You can have the coolest, most hyper-intelligent system ever devised, but that cannot see patterns which do not exist. Of course market data has patterns (sell on Friday, for instance), but it is always limited. Generally, I think you could even in principle never get more than 80% accuracy with any kind of neural network, because of the lack of complete patterns in market data, whether we are talking about a rat or a human or a whatever kind of neural network. Which brings me to a second point, since 100% accuracy or anywhere near that is not possible even in principle, applying electric shocks to test subjects when they are wrong is simply cruel and irresponsible.
"Since I did not want to make it too complicated, I only used the USD/EUR future to make the rats experts in this specific market segment, but other rats can be trained in other markets as well. I trained the rats for about 3 months, starting with 80 Sprague Dawley laboratory rats, 40 males and 40 females with the intention to cross the best of them to genetically create the best traders through select breeding."
Sounds like a pigeonrank algorithm to me...