You are using an out of date browser. It may not display this or other websites correctly.
You should upgrade or use an alternative browser.
You should upgrade or use an alternative browser.
AI Any thoughts
- Thread starter Bogan7
- Start date
Historically, very few people have had much luck with neural software in predicting market behavior.
Recently however, I read an article in Wired magazine about Jeff Hawkins who started Palm. He's developed a new type of software that mimics how the human brain works called Numenta. Maybe this is the type of breakthrough needed to get results in this area.
Runningbear
Recently however, I read an article in Wired magazine about Jeff Hawkins who started Palm. He's developed a new type of software that mimics how the human brain works called Numenta. Maybe this is the type of breakthrough needed to get results in this area.
Runningbear
According to wikipedia neural AI networks have shown promise.
Interesting that you mention Numenta. I just finished reading Jeff Hawkin's book "On Intelligence" which explains his theory of intelligence and the principles behind Hierarchical Temporal Memory (aka HTM, Numenta's technology). He eschews the term "artifical intelligence" given AI's failure to answer what intelligence actually is, and attributes the fall of artificial neural networks to that reason as well. Instead, HTM is an algorithmic model of neocortical memory, based on 4 axioms:Quote from Runningbear:
Historically, very few people have had much luck with neural software in predicting market behavior.
Recently however, I read an article in Wired magazine about Jeff Hawkins who started Palm. He's developed a new type of software that mimics how the human brain works called Numenta. Maybe this is the type of breakthrough needed to get results in this area.
Runningbear
- The neocortex stores sequences of patterns
- The neocortex uses invariant representations of patterns
- The neocortex recalls patterns auto-associatively
- The neocortex stores patterns in a tree-shaped hierarchy
Those four elements enable brains to make predictions. Correlations between spatial-temporal patterns is essentially how Hierarchical Temporal Memory works... with a different hierarchical structure for the problem domain. I use a similar approach to analyze sizes changes on the order book mapped to price change.
I've made some money using these AI forecasts: http://www.understandmarket.com/index.php?categoryid=23
But I'd say that a trader must have his head in the right place to use them.
But I'd say that a trader must have his head in the right place to use them.
From the linked site in your post: "It is based on analysis of pure price rather than information distorted by technical indicators."Quote from UM_manager:
I've made some money using these AI forecasts: http://www.understandmarket.com/index.php?categoryid=23
But I'd say that a trader must have his head in the right place to use them.
So what's the difference? The objective probability is still 50/50. If anything, it's a complete misuse of neural networks.
Quote from Joe Doaks:
I once had a fairly accurate (and probably totally spurious) linear estimator of the S&P 500 based on the yield, the 90 day T-Bill rate, and M3 (you remember what THAT was, haha!). I got there by correlating the S&P with every bit of data I could find, like the size of the Ukrainian potato harvest and the price of rubbers at the corner gas station. Only the aforementioned three had convincingly high correlation coefficients.
Think about this for a minute:
A neuron has no knowledge of causality. It has to be presented with high-level sensory data to learn about its world. More importantly, it must be a spatial-temportal pattern occuring contiguously in time.
e.g., As I type this, I am using a multitude of senses: touch carried by pressure against my fingers, hearing the keys being pressed, seeing the characters on the screen, etc. If just one of these sensory inputs were removed, I wouldn't be able to elucidate this.
In the context of finanical markets, sensory inputs would be datasets that are correlated in space and time. i.e., Nothing can be derived from a time series of one currency pair (contrary to that site).
As you know, there are many forces that influence currencies... oil prices, interest rates, economic data, and unquantifiable forces such as the overall business environment. Those would be some of your inputs.
Who can argue with that? However, what happen if Neural Networks would be trained not only on indicator values but and with those additional data like oil prices, economic data and so on? And if we take into account that computational capability of NN is not limited with only 5-7 variables like we are, then result would be very impressive.