United States
Member #83,698
December 13, 2009
225 Posts
Offline
Quote: Originally posted by jimmy4164 on Aug 27, 2010
Made the mistake of peaking after logout!
My High/Low predictor model uses the (Open minus Yesterday's Close), so I KNOW the Open.
--Later
Sorry, I thought you were trying to predict the open. The day's open minus yesterday's close is an interesting index as it's an indicator what the latent demand is. The market maker would've selected the open such that it would sweep through any orders, stops and limits placed so as to maximize his brokerage's commissions. This would give you the initial day's momentum which may very well define what the day's high or low would be. Another interesting indicator to factor in would be the product of the previous day's average minus close (I was going to say open minus close but as I've mentioned, the open has an artificial element hence perhaps an average would be more appropriate) and the volume to see if there's an overall consumer based momentum.
It's kind of interesting that some writer's such as Alex Doulis ("Take Your Money and Run") say that the stock market has become too efficient and that you profit best from inefficiencies hence bond markets and hedge funds are more attractive. Yet the efficiencies are also making it more possible to harvest volatility as per Shannon et al.
United States
Member #93,943
July 10, 2010
2,180 Posts
Offline
Quote: Originally posted by jwhou on Aug 27, 2010
Sorry, I thought you were trying to predict the open. The day's open minus yesterday's close is an interesting index as it's an indicator what the latent demand is. The market maker would've selected the open such that it would sweep through any orders, stops and limits placed so as to maximize his brokerage's commissions. This would give you the initial day's momentum which may very well define what the day's high or low would be. Another interesting indicator to factor in would be the product of the previous day's average minus close (I was going to say open minus close but as I've mentioned, the open has an artificial element hence perhaps an average would be more appropriate) and the volume to see if there's an overall consumer based momentum.
It's kind of interesting that some writer's such as Alex Doulis ("Take Your Money and Run") say that the stock market has become too efficient and that you profit best from inefficiencies hence bond markets and hedge funds are more attractive. Yet the efficiencies are also making it more possible to harvest volatility as per Shannon et al.
"Another interesting indicator to factor in would be the product of the previous day's average minus close (I was going to say open minus close but as I've mentioned, the open has an artificial element hence perhaps an average would be more appropriate) and the volume to see if there's an overall consumer based momentum."
Interesting idea to avoid the Open ambiguity. When I'm optimizing 2 or more parameters I assume the network software is sorting out their products with at least one hidden layer. I look at V, [delta]V, C, and [delta]C, among other things, so if I include (Avg - C), Backpropagation should find the meaningful products. (Right?)
BTW, does this help with your system of equations?
United States
Member #83,698
December 13, 2009
225 Posts
Offline
Quote: Originally posted by jimmy4164 on Aug 27, 2010
"Another interesting indicator to factor in would be the product of the previous day's average minus close (I was going to say open minus close but as I've mentioned, the open has an artificial element hence perhaps an average would be more appropriate) and the volume to see if there's an overall consumer based momentum."
Interesting idea to avoid the Open ambiguity. When I'm optimizing 2 or more parameters I assume the network software is sorting out their products with at least one hidden layer. I look at V, [delta]V, C, and [delta]C, among other things, so if I include (Avg - C), Backpropagation should find the meaningful products. (Right?)
BTW, does this help with your system of equations?
That's the interesting thing about neural nets. We never really know what it finds out. At various times, people would try to work backwards and try to determine what logic the neural net has developed but of course, it could be anything.
With regards to non-linear programming. I was really hoping to avoid doing a 38 dimensional gradient for a min max Newton method solution. Just the very mention of 38 dimensions turns my head into a pretzel. It may all be moot anyways, although the method I was thinking of is mathematically sound, it still lacks sufficient causality for it to be useful. It was an interesting thought experiment (It was an attempt to read Bingo style scratchers without scratching by treating it as a cryptographic simple substitution code). The results are as I expected, the cards read as possible loser and possible winner, it would have them at various probability levels but without a causality link, those probabilities are meaningless. I'll post a summary of the experiment in the scratchers section entitled Enigma machine for Scratchers.
United States
Member #93,943
July 10, 2010
2,180 Posts
Offline
Quote: Originally posted by jwhou on Aug 27, 2010
That's the interesting thing about neural nets. We never really know what it finds out. At various times, people would try to work backwards and try to determine what logic the neural net has developed but of course, it could be anything.
With regards to non-linear programming. I was really hoping to avoid doing a 38 dimensional gradient for a min max Newton method solution. Just the very mention of 38 dimensions turns my head into a pretzel. It may all be moot anyways, although the method I was thinking of is mathematically sound, it still lacks sufficient causality for it to be useful. It was an interesting thought experiment (It was an attempt to read Bingo style scratchers without scratching by treating it as a cryptographic simple substitution code). The results are as I expected, the cards read as possible loser and possible winner, it would have them at various probability levels but without a causality link, those probabilities are meaningless. I'll post a summary of the experiment in the scratchers section entitled Enigma machine for Scratchers.
But I'm still wondering about this question, which we both forgot about last night!
"...but really would like you to address the issue of the reasonableness of Backpropagation of essentially random inputs, and overtraining..."
United States
Member #83,698
December 13, 2009
225 Posts
Offline
Quote: Originally posted by jimmy4164 on Aug 27, 2010
But I'm still wondering about this question, which we both forgot about last night!
"...but really would like you to address the issue of the reasonableness of Backpropagation of essentially random inputs, and overtraining..."
I would say that you have to assume that what has been learned from previous iterations had been of some value so the danger is in having that value negated by noise in subsequent iterations. Of course, noise should cancel out but that isn't a given. The assumption that previous back propagation iterations have value that you would not want to lose leads to the concept that each iteration should adjust the values by a fraction of the previous iteration. Intuitively I would suggest that the increments would be an inverse exponential relationship to the number of iterations. This would reflect the concept of an initial steep learning curve and an eventual plateau. With a case where there are hidden layers, perhaps the assertion of an additional element in the hidden layer should be complemented with resetting the backpropagation increments to a certain value, perhaps 0.75 or 0.5. If you have a random mutation effect in the hidden layers where there's a small chance with each iteration that a new concept may be introduced then it would be logical for the increments to reset when such a mutation occurs.
There may be a rationale of backpropagation of known random inputs and that is to try and average out the effect of any random components from actual training. In this case the backpropagation increments should probably be a fixed value such that a standard deviation in number of random increments in one direction not exceed the scale of the increments that you used with real data since again you don't want the randomness to overwhelm any real relationships developed. Introducing known random noise may allow the neural net to literally "think outside the box" and pursue another local minimum/maximum solution. I would prefer the occasional random addition of an asserted element in the hidden layers over feeding intentional random inputs.
United States
Member #93,943
July 10, 2010
2,180 Posts
Offline
Quote: Originally posted by jwhou on Aug 28, 2010
I would say that you have to assume that what has been learned from previous iterations had been of some value so the danger is in having that value negated by noise in subsequent iterations. Of course, noise should cancel out but that isn't a given. The assumption that previous back propagation iterations have value that you would not want to lose leads to the concept that each iteration should adjust the values by a fraction of the previous iteration. Intuitively I would suggest that the increments would be an inverse exponential relationship to the number of iterations. This would reflect the concept of an initial steep learning curve and an eventual plateau. With a case where there are hidden layers, perhaps the assertion of an additional element in the hidden layer should be complemented with resetting the backpropagation increments to a certain value, perhaps 0.75 or 0.5. If you have a random mutation effect in the hidden layers where there's a small chance with each iteration that a new concept may be introduced then it would be logical for the increments to reset when such a mutation occurs.
There may be a rationale of backpropagation of known random inputs and that is to try and average out the effect of any random components from actual training. In this case the backpropagation increments should probably be a fixed value such that a standard deviation in number of random increments in one direction not exceed the scale of the increments that you used with real data since again you don't want the randomness to overwhelm any real relationships developed. Introducing known random noise may allow the neural net to literally "think outside the box" and pursue another local minimum/maximum solution. I would prefer the occasional random addition of an asserted element in the hidden layers over feeding intentional random inputs.
This doesn't agree with my experiences trying to find predictive value in data with at best marginal non-random elements. In such cases, to elicit any useful information from the inputs required reducing the number of variables and/or hidden layers, rather than increasing them. Otherwise, the result was "curve fitting" extrordinaire!
"There may be a rationale of backpropagation of known random inputs and that is to try and average out the effect of any random components from actual training."
This statement implies that the sought after, non-random elements are the rule, rather than the exception. In the current case under inspection, based on my analysis, I must disagree.
United States
Member #83,698
December 13, 2009
225 Posts
Offline
Quote: Originally posted by jimmy4164 on Aug 28, 2010
This doesn't agree with my experiences trying to find predictive value in data with at best marginal non-random elements. In such cases, to elicit any useful information from the inputs required reducing the number of variables and/or hidden layers, rather than increasing them. Otherwise, the result was "curve fitting" extrordinaire!
"There may be a rationale of backpropagation of known random inputs and that is to try and average out the effect of any random components from actual training."
This statement implies that the sought after, non-random elements are the rule, rather than the exception. In the current case under inspection, based on my analysis, I must disagree.
Well, I wasn't thinking of increasing the number of hidden layers but the very infrequent additional hidden layer element being asserted which has the effect of allowing the neural net to consider another avenue of thought. Of course such an assertion has a destabilizing effect and an excessive number of exerted elements in the hidden layer would be excessive curve fitting. Ideally, you want the minimum number of elements in the hidden layer exerted. I always wondered about that since many three layer neural nets have 50% of the hidden layer elements asserted by default.
If the noise to signal ratio is too high, a neural net may not be able to help, you'll need to have some non-random events to look for.
How goes the day trading? Do you hedge your bets? Do they allow shorting warrants in the States? Without warrants, how would you hedge? If you don't negate volatility with a hedge, are you trying to harvest volatility? What's your strategy for harvesting volatility? How do you minimize your transaction costs?
United States
Member #93,943
July 10, 2010
2,180 Posts
Offline
Quote: Originally posted by jwhou on Aug 28, 2010
Well, I wasn't thinking of increasing the number of hidden layers but the very infrequent additional hidden layer element being asserted which has the effect of allowing the neural net to consider another avenue of thought. Of course such an assertion has a destabilizing effect and an excessive number of exerted elements in the hidden layer would be excessive curve fitting. Ideally, you want the minimum number of elements in the hidden layer exerted. I always wondered about that since many three layer neural nets have 50% of the hidden layer elements asserted by default.
If the noise to signal ratio is too high, a neural net may not be able to help, you'll need to have some non-random events to look for.
How goes the day trading? Do you hedge your bets? Do they allow shorting warrants in the States? Without warrants, how would you hedge? If you don't negate volatility with a hedge, are you trying to harvest volatility? What's your strategy for harvesting volatility? How do you minimize your transaction costs?
In the case of inputs from lottery results, I believe the noise to signal ratio is VERY high. TOO high!
For daytrading, my vehicle of choice is QQQQ for Long and PSQ for short. I use a simple approach, no warrants, futures, or options. I'll occasionaly pair PSQ and QQQQ as a short term hedge, but generally, I rely on close scrutiny of my [one] trade, my indicators, and the "kill" switch! Even Kelley's Criterion allows for quite liberal percentages of capital in this market.
I'm an adherent of Trading in the Zone by Mark Douglas, and Design, Testing, and Optimization of Trading Systems by Robert Pardo.
Tahiti- Polynesia Tuvalu
Member #34,523
March 4, 2006
57 Posts
Offline
Hi,
I wonder how Lotto can be defined. Is it a chaotic or a random or both space state vector? Is it linear or nonlinear, stationary or not stationary, deterministic or not, periodic or not periodic? Would the definition change if I use the drawing in grawing order or sorting order?
It is probably chaotic as the lyapunov exponent and shannon entropy are positive. I don't know for others parameters.
Defining the caracteristics of each lotto is important in choosing the right tools for analysis.
United States
Member #93,943
July 10, 2010
2,180 Posts
Offline
Quote: Originally posted by bob790 on Sep 17, 2010
Hi,
I wonder how Lotto can be defined. Is it a chaotic or a random or both space state vector? Is it linear or nonlinear, stationary or not stationary, deterministic or not, periodic or not periodic? Would the definition change if I use the drawing in grawing order or sorting order?
It is probably chaotic as the lyapunov exponent and shannon entropy are positive. I don't know for others parameters.
Defining the caracteristics of each lotto is important in choosing the right tools for analysis.
Anyone knows about that subject?
Thank you very much.
bob790,
Most of your questions above are outside my areas of expertise. I am familiar with Shannon's analyses as they apply to changes in stock prices. However, believing that in the present, the probability of a past event is 1.0, I fail to see how his work would have much value applied to the classes of chaos/randomness produced by lottery ping pong ball machines or random number generators producing sequences with extremely long periods.
Edinburgh United Kingdom
Member #97,829
September 24, 2010
63 Posts
Offline
Hi jimmy4164,
Thanks for your reply. I haven't made any changes to my browser and/or security software. I do not know why I can't log in as martor54. I follow the procedure, I get "Log in successful" but my user name is not there. However, I can do that "no problemo" under my slightly changed user name. I'll have to stick with that.