I hope this simple analogy doesn't 'splode your brain
I looked briefly on the net for one, but couldn't find one simple enough, imo.
There have been some great recent posts here regarding April 15 statistics going back to, let's say 1955 or so. Typical statistics will add all the percentage changes for April 15 (or for where April 15th falls on a Tuesday, etc. etc.) for all those years and achieve a bias for that day. An average. I think Ebb does something similar.
Someone willing to put in a lot more work could work out the averages for each year,
individually, from 1955-2013, analyse the data, and possibly determine that the bias varies by a predictable amount each year, and then formulate an algorithm or formula to predict the bias
as it relates to each year. Such a person would have an advantage over someone simply using a simple average over all the years to predict a 2014 bias for April 15, given only prior years data without an algorithm explaining how the bias is changing over the years.
Suppose differently that someone looks at the averages based upon what day of the week April 15 falls. Like above, this can easily be done with simple statistics. They would conclude to the effect of: bias when it falls on a Monday is this....on a Tuesday is that, or the other....
Now suppose they wanted to know the bias when the previous close was 'up' versus 'down,' AND when the close before that was up or down; OR instead of just up or down; when its up 0-0.5%, versus 0.5-1.0%, versus ....., 0-0.1%, (as the percentage divisions approach infinity and the size of each division approaches zero). A realistically impossible task for a human.
Humans typically only compare one or two inputs at a time in these sort of cases: April 15th bias, April 15th and Monday bias, etc.
A NN, like above, could factor in the year as well as the day of the week as well as infinite price 'divisions' as well as the closes of the last several days as well as ...lunar cycles... etc. A NN running on a typical home computer can effectively handle scores of inputs, with thousands of samples/examples to analysis. Virtually impossible for a human with Excel or even an advance trading platform to accomplish (unless it properly utilizes some extremely customizable machine learning technology).
A NN is statistics on steroids...
I could go on: Why they don't "work" for some folks. Why they do work for some folks. Etc.
Hope this helps, if not, I can try again if you like.
IMO,
Q
Bookmarks