Can anyone give me any idea about neural network trading?

<p id="isPasted">Everything looks easy and solvable until you stumble over the problem that having enough data AND knowing how to do the groundwork (cleaning it up; tossing away what is noise and keeping what it is important to make a “not so bad”-prediction) is the real show-stopper.</p><p>Machine Learning/AI/Neural Networks/Deep Learning/whatever looks supereasy in every tutorial. Because it comes with prepared, cleaned datasets that are readily available. In the real world you need to gather and prepare this data first.</p><p>The problem is the data. Where does it comes from? What the hell am I supposed to do with it? Why the hell is it so incomplete? Do I have the domain knowledge to know what to look for? Who has it? Does it come for free? Oh snap!</p>

<p>At best neural networks in some instances will give those trading such technique a small edge, which in general is the best you can do re: trading techniques. That’s not to say you can’t make lots of money with small edges; you can, and that’s exactly why/how casinos virtually print money (well, ignoring cost of operations and mismanagement). I refer you to the information on the wizardofodds web site for detailed statistical information on the power of many small “bets” with a small (1–3%) edge. It is very, very striking.</p>

<p id="isPasted">I won't talk specifically about Neural Networks, but more generally about expert systems and algorithmic trading. The major players in the market do study these techniques extensively.</p><p>The major problem is that there is no technical analysis that consistently yields returns. And in fact, there can't be. As soon as a method of identifying good trades is discovered and put into use, it becomes useless.</p>

<p id="isPasted">Individual forex trading is largely a game of technical analysis and intuition building. At the levels of leverage required to make good money, you can’t hold positions long enough for most fundamental changes to impact your trade.</p><p>As such, what you would be doing is essentially pattern recognition across time, likely using a shallow neural network or maybe go the other way and start doing LSTM / RNN type deep learning. The problem with this is that you would be training the algorithm on past data without much context.</p><p>If your plan is to train the network on technical signals like candlestick patterns, then I would advise against it. A simpler method would suffice while being much more explainable and optimizable. If on the other hand, you are talking about discovering new patterns? Then I think there are many possibilities, but it will definitely cost you quite a lot to test out your theorems.</p>

<p id="isPasted">It’s not about the tool; it’s about the market. Let me talk about two things.</p><p>Firstly. The stock market (and practically anything financial) is an overwhelming amount of noise with some faint signal hidden somewhere when we’re lucky. Nothing can find more information in something than what there is but, and that’s why neural networks are a bit tricky, the more complex your model, the more spurious signals it can find.</p><p>You know how we can see something into practically any of those cute fluffy clouds on a sunny day? Neural networks are at least as good as that finding patterns in anything.</p><p>I’m not saying it’s impossible to use NNs in the context of the stock market but it’s probably one of the trickier applications.</p><p>Secondly. The stock market (and practically anything financial) is not predictable, period. Now, I won’t deny that it’s somewhat predictable most of the time. However, the probabilities of unexpected extreme events are large enough that getting most of the smaller ones will still not protect you from getting bankrupt by that one unexpected large event.</p><p>Basically, if you depend on prediction alone, you’ll go bust. You gotta make sure that your exposure is such that no unexpected event can hurt you more than what you can survive (yes, it will cost you, but less than going bust). Once that’s done, you can start thinking about prediction.</p>