I am not a neural network expert but have thought about using neural network in developing trading strategies for quite some time. in the end, I found it is almost an impossible task in terms of using times series data as input. the reasoning is quite simple. for example, suppose there are 100 ways to make a profit in the market, each of the 100 strategies corresponds to a specific context of the data points, e.g., double bottom, shoulder and header, breakout, pin bar, etc. if you input all the data points to the neural network, the NN system can not pick out any of the 100 strategies, since every strategy corresponds to one pattern, i.e., one specific set of parameters in NN, thus there are no parameter sets in NN which can distinguish those 100 patterns. if NN can not produce a set of parameters to categorize specifically each of those 100 patterns, they can not pick out any of the 100 strategies either. for the above problem, genetic algorithm would be much easier to apply. NN is great if you have specified the search domain. any opinions?
1) Think about what you're trying to predict. A point prediction is likely to just identify the general trend. It's also a bit similar to what you describe in OP. 2) Trading on a prediction may be more involved than "if price > predicted; short, else long" due to transaction costs, risk management and other reasons. 3) Neural networks are great for overfitting (this is probably the most common complaint) on past data with poor out of sample performance; however this probably much relates to incorrect usage also. 4) Lack of training data for some bar sizes (e.g. daily) can be an issue. Some people (also on this forum) claim to use NNs for trading, but the way they are using them is more involved than predicting what the price will be one hour from now or whether a picture is of a cat (as you would typically learn on an introductory deep learning course). Do note that while NNs are hyped they are just one of many ML algorithms. In particular, check whether your problem can be solved with linear/logistic regression, random forests, or K nearest neighbor classification first - the more simple the more likely to work out of sample. Either way, garbage in garbage out principle applies and what you feed into your ML algo matters more than algo itself.
Making money in trading is much more than predicting the price/trend. And why are you referring to only one NN? If they are 100 ways to make money (like you said), why not have 100 different NNs? Having said that, I started to use NN 25 years ago, along with genetic algo, and NN of genetic algo. They have their specific usages. But making money is not one of them (IMO). They're only one part of the solution. Again, making money in trading is much more than just predicting price/trend. We're not talking about lottery here where you make money if you pick the right numbers. The best way to learn NN is to use them and play with them. There will be people saying that you cannot make money with NN, and there will people saying the reverse. Will that help you? Please re-read the first sentence of my post. ps: My experience will not help you making money with trading.
I am a big fan of random forests but my results depend highly on how I flow the data through. I am working through some deep learning to see if I can improve things but I don't know if it'll be worth it. I thought I'd need deep learning to identify post-hoc trends automatically but turns out it wasn't necessary.
You can't just feed a NN a time series data and expect it to pop out a trade able strategy. 70% of your time should be building a data set with decent predictors. Next you will most likely have to transform the Dependent Variable due to time varying volatility, non-normality etc.. Your model is only as good as the inputs. "Garbage in garbage out" Also, ask yourself - why you are using a neural net over xgboost or random forest? Each model cuts up the data a bit differently. Here is a picture of how different models look at the data. PS. FeatureTools package in Python might be able to help you with automated feature engineering. It's under development, but it looks promising! Maybe someone with experience using it can add something.
I worked with wealth labs neural network extension 10 years ago. After 3 days of training it learned to buy every downtick and sell every uptick
No need for a thesis...just simulate trade it and then post your results anywhere online...next...apply it with real money. wrbtrader
ok - the NN is simply not good enough! look up AlphaZero - tell the network the basic rules, it figures out chess and go in no time.. the rule is simple in trading - the biggest profit wins! DeepMind is doing some amazing stuff, they are just not into running a hedge fund, yet.