QuoteOriginally posted by: KackToodlesQuoteOriginally posted by: twofishOne application of NN in which the inability to figure out what the neural net did is actually useful is spam filtering. If you have a transparent algorithm filtering spam, then the spammers can easily counter the filter, but the fact that it is hard for the NN to describe why an e-mail was marked as spam makes it harder for spammers to counteract the filter.Can NN's pick stocks sucessfully on a consistent risk-adjusted basis?NNs are what's called 'universal approximators'. Fuzzy logic is similar. If you know how to properly train them (which is a big IF) they will detect and quantify all causal relationships. NN R&D (may years ago I was into that at DEC) requires a very strong backgound in differential geometry and that level of math is needed to get almost anything useful working properly.

Neural nets are one of those interesting things that a lot of people pick up on at the start of a PhD and play about with, but rarely get anything useful out of.Therefore there are a lot of people about who can claim experience, but few who can provide evidence of having really usefully utilised them.Not therefore the best way of making your CV stand out.Prof N is no doubt an exception.

Not therefore the best way of making your CV stand out.Prof N is no doubt an exception.CV? Oh, you mean the thing I haven't updated in 20 years...

There is an example: S(0) =2, K=2, r = 0. And S(1) = 4 with p =0.01 and S(1) = 1 with p=0.99 lead to the same call option price as S(1) = 4 with p =0.99 and S(1) = 1 with p=0.01. In both cases the volatility is the same. This example is a sufficient condition to say that option price and its interpretation with risk neutral probabilities is mindless.

I could be wrong but I I think the fund being run by Olsen (www.olsen.ch) is using some kind of crazy NN, combined with some of the ideas espoused by Mandelbrot, i.e. that fractal patterns exist in financial time series. If anyone manages to find out what they're doing id love to know.Regards,Ren.

QuoteOriginally posted by: CuchulainnQuoteAnothe application: you produce a wider mesh of results and "interpolate" by ml methods (take in account that the simulation of a complex process behaviour, with sytems of over 80 PDEs, often takes hours and cannot be done in "real time")Wow! The number is increasing by the day How does such a PDE system look like? Are you talking about 80 factors or a low-factor problem with a vector solution?Yesterday someone wanted to solve an 11-factor PRDC system.I am not the maker, therefore I can only give a poor generalist's answer. No it is not 80 factor with full dependence of all factors.But if you want a complete 3D model of a blast furnace process you need to model transitions of ore, coke,...to liquid iron from top to bottom; gas, from bottom to top. You model reactions, aggregate state transitions, convection, diffusion and have systems for material balances, energy balances,... Clear, not all factors interrelate but you have enough depenencies to call this A system (of 80 PDEs).To solve it is not so easy (FE is the core method but you need "tricks" to improve it).But in relation: to achieve the same accuracy and speed of the analysis as it is required in QF (say, a few basis points of temperature or C, Si, Mn predictions) the system needed to be simplified significantly.

Off the top of my head I'd guess that the typical supernova simulation is a system of about a dozen coupled non-linear PDE's. If you add a r-process reaction network to it, you are talking about hundreds of factors. (You have a factor for each isotope, and since these are not in nuclear statistical equilibrium, you have to figure out the reaction rates for each transition. Nasty.)The cool thing about all of this is when you look at all of these equations and try to figure out the essential "ghost" that describes the basic behavior.

QuoteOriginally posted by: twofishOff the top of my head I'd guess that the typical supernova simulation is a system of about a dozen coupled non-linear PDE's. If you add a r-process reaction network to it, you are talking about hundreds of factors. (You have a factor for each isotope, and since these are not in nuclear statistical equilibrium, you have to figure out the reaction rates for each transition. Nasty.)The cool thing about all of this is when you look at all of these equations and try to figure out the essential "ghost" that describes the basic behavior.Aren't we talking about n-body problems here, where you define the interaction of single particles and then take limits as N -> infinity ? I think this is commonly done in neuro-biology, agent-based modelling, galaxy cluster modelling, etc

you might google adam beger's MaxEnt tutorial, available online when I was an undergrad in the late nineties. MaxEnt is the cousin to NN (NN with a sigmoid function) and here is a technique of great versitility with definite application to finance. I use it in the context of modelling in physics, particularly with a time series with truncation problems; MaxEnt is simply the best backward/forward predictor and a great (supervised) ML tool for your bag of tricks

Here you gocorrection, the guy's name is berger

QuoteOriginally posted by: zetaHere you gocorrection, the guy's name is bergerdiscrete-time... yuck, I don't think the MaxEnt is the be-all end-all.good stuff about information theory, there is something deeper at work here because I've seen similiar w/ regard to the exponential family of distributions being some of the only that can be solved exactly in closed-form for diffusion inference.

It ain't pretty, but it will get you out of a jam. There's nothing so information theoretic as the idea that when you 'know nothing', the distribution should be as uniform as can be ie have maximal entropy. Depending on the implementation, it is pretty fast too. I think Csiszar gives example in 'maxent, mathematics and information theory' found in 'Maximum Entropy and Bayesian Methods' , Kluwer, 1996

Isn't "entropy ranking" used to create ID3 decision trees?Fuzzyfied ID3 decision trees are nice; they are readable and computational (predictive).BTW, talking about ml methods. To undestand whether your vast multi-dimensional data are correlated in a way, you might get a first view by applying SOMs (self organizimg maps, with Kohonen's "the-winner-gets-it-all" algorithm (with ideas from NN theory)). It is for unsupervosed learning, or call it data visualisation.Introduction: SOM

QuoteOriginally posted by: exneratunriskIsn't "entropy ranking" used to create ID3 decision trees?Fuzzyfied ID3 decision trees are nice; they are readable and computational (predictive).BTW, talking about ml methods. To undestand whether your vast multi-dimensional data are correlated in a way, you might get a first view by applying SOMs (self organizimg maps, with Kohonen's "the-winner-gets-it-all" algorithm (with ideas from NN theory)). It is for unsupervosed learning, or call it data visualisation.Introduction: SOMI wouldn't bother about SOM, it has been superceded by the Generative Topographic Mapping

QuoteOriginally posted by: crowlogicI wouldn't bother about SOM, it has been superceded by the Generative Topographic Mappingthere are several approaches to generalize, extend or combine methods.I am biased! Because, this is from partners of mine:SOMs, Clustering, ml