Regularization is some means of complexity control. You switch form trying to find the model that has minimal error on your data to finding the model that minimizes "error(model, data) + C*complexity(model)". There are many ways to measure the complexity, and usually the value of C is pretty unimportante (it is empirically found that it only affects the solutions when you change it in orders of magnitude). Keeping a test set is always testing your predictions on held-out data that you didn't use to train the model and that is hopefully independent from your training data.
Still, if you come up with a huge number of models and test them on the same test data you're statistically bound to find a few that seem to work very well, until you try it in different data.
My feeling with most of these market-beating schemes is that just the price history of stocks shouldn't be that predictive of future prices (but aggregating real-world information should). After all, usually what creates spikes and valleys in the prices are some relevant news or decisions.
To wit, P(model|data)=P(data|model)P(model) / P(data), which is to say that you calculate the probability a model is true given the data you see based on the probability you see the data you see given a certain model, multiplied by the probability of the model you are using.
I guess the problem in all this is we want* the model to be very complex, but we can't tell that complexity from overfitting (and it may be that no model exists so the only thing you can do is overfit.)
Being bayesian gives you an extra capability that is model averaging, and this does usually improve the behavior at a high computational cost.
I really like bayesian models, and right now I'm experimenting with one that should do unsupervised sentiment analysis without a priori knowledge of word polarity or things like that (yes, I'm a phd student in machine learning).
It also goes the other way, and some formulas are first proposed in a more bayesian-ish context and then extended to some simpler-looking empirical formulas (for example, the jumps from hidden Markov models to max-ent Markov models to conditional random fields to max-margin Markov networks to structured SVMs).
There are more approaches to machine learning, and in John Langford's blog there is a nice table showing the merits and flaws of many of them: http://hunch.net/?p=224 . But you must keep in mind that you can find many equivalencies between these approaches (boosting for example can be seen as a loss minimization with regularization, and max-ent can be seen as a special case of a bayesian model, etc).
People project patterns onto data when in fact, they don't exists.
Edited to add this, which is sometimes also brought up: http://en.wikipedia.org/wiki/Illusion_of_control
So, the patterns we see in the Ulam spiral can be completely random?
We observe patterns. Human brains are really good in finding them. The problem is finding patterns in past data may or may not provide useful insights into future data.
With the resources being dedicated to this problem, I find it highly unlikely I could be able to offer any helpful advice that would prove more helpful than what already exists.
It would be hard enough to predict the market if the only factors influencing it were signals coming out of the companies we trade. Predicting what other algos may me doing in the space we are trying to examine and the influence their trades exert in its future is... Well... I won't try.
I'm not extremely well educated on the subject, so somebody correct me, but these predictive models have always struck me as just complicated splines. Add a thousand data points, and the "predictions" for old datapoints look really good. The problem isn't with the math, it's with an inappropriate application of the math based either on fundamental misunderstanding of extrapolation or on a fundamental failing of human psychology in not realizing that the model was built on the data, and not the other way around.