What I object to is the argument that using 5 years of stock data works better for future system performance rather than 1 year worth of stock data.
I might not have been too clear in my post. I didn't mean to imply that 5 years of data works better than 1 year. What I was saying is that you can't simply use 1 year's worth of data, find the best parameters (i.e. the Model) and then use it as the definitive Model (over 5 years, 10 years, 3 months, or whatever).
Rather you need to create a Model based on a large number of data points as well as various time periods that encompass different market conditions. Only after you have a Model that works well over all these periods can you be reasonably sure it will do a good job in the future.
Your statement might be based on your experience that for many companies the 5-year track record is more revealing for the future than the last year.
So no, for the reasons just given above, I don't believe the 5 year track record is more revealing than the 1 year record in the general case.
Regarding curve fitting, after re-reading your posts, I think I am probably talking about something different from what you're talking about.
If we have a number of data points (e.g. obtained from a survey), we can try to fit a line (or curve) through these points. We can then obtain the function for this line by filling in the appropriate coefficients. Perhaps that is what you were referring to.
I was combining this thought with that of optimization.
If we have a set of data points (in this case share prices and dates) and we also have a function (i.e. the AI algorithm) then we want to find the "coefficients" (i.e. AI parameters) that will give us the best returns. To do this we fit a curve (actually a 3-dimensional area, Return on z-axis, date on x-axis and price on y-axis) through these points. We then obtain the function for this curve by filling in the coefficients using something like the method of least squares.
Now if we apply this function to a new set of data points (i.e. a new set of dates and prices), it might predict the best returns if we used a sufficiently large enough (and well chosen) set of original data points. However if we use a poorly chosen set of data points, this function is overfitted to those points and will not do a good job of predicting returns in the future.
In effect it is approaching the optimization problem from the opposite direction.
Having said this, I've never actually considered doing this for real as I prefer to use brute-force optimization (and soon I'll be using a genetic algorithm once I have the time to implement it). So I don't know if it will work in practice or not.
"Do a curve fit on a saw tooth wave and my point will be clear to all....you get the profile of a very worn out saw."
My experience indicates all waves are saw toothed. I'm not implying the points maintain an order or structure but just that it's possible to track the swings without a curve. There's an expected velocity associated with every movement. Actually minimum and maximum expectations. The velocity is either there or it isn't and when it isn't and it's confirmed by price moving in the opposite direction then it's a trend reversal. A single bar moving the opposite direction doesn't mean the trend is changing but it is an alternate direction swing. As long as contra-trend swings are confined within the current moves normal velocity the trend is making progress.
Conrad said "Do a curve fit on a saw tooth wave and my point will be clear to all....you get the profile of a very worn out saw."
I definitely agree.
Not quite the sine wave I pictured but it uses the preferred average(14). I didn't spend much time and Excel isn't my favorite platform but given time and a higher research budget <g> I might hit the right combo. It's 90 degrees out of phase rather than the anticipated 180.
Market Data 
Markets 
AI
