well my nerual nets converge so fast im considering trying to befriend them. if you can't beat them join them. however at first i had to do a lot of work they dont thank you for it.
the esma needs to add balances can not go below -10 euros surely or else you would not get analysis you would get nonsense
There are significant issues using Neural Networks for time series modeling (and trading) Issues: You need to have a sparse nn architecture (i.e. not have fully connected layers). This is because you cannot have future data impacting the weights / biases of older data. Unfortunately, you literally cannot use PyTorch if you want to do proper time series modeling because the sequential mask cannot handle sparse architecture (a silly design choice by the PyTorch team). Tensorflow also has pretty convoluted and messy ways of handling sparse matrix algebra. Time series modeling should be more concerned with looking at historical data (look at lagged series). In a nn this would require INTRA-layer connections as opposed to (or in addition to) INTER-layer connections. Intra-layer connections would provide directionality to the model, while it is not really present in traditional network architectures. For financial data, you have to transform the data before you feed it in. As previous posters said: garbage in, garbage out. For example, people typically look at the log returns of stock prices because it has nice mathematical properties. You have to remember that these models are optimization models. They are there to optimize the parameters, not to predict the future from seemingly random financial data. Solution: The only solution is to not use neural networks for this problem. The only time series applications that can work well with neural networks are problems involving stationary data. For example, weather data and unemployment rate data are stationary (cyclical), so a nn model would be extremely effective in finding an optimal model. If you really want to model market dynamics, I suggest looking into learning about stochastic calculus, measure theory, and real analysis.
You are right, not only are you not an expert in his fields but you hardly know anything. Most of the claims you made are false and your reasoning is incoherent. Without going into details, you are simply wrong. I work in this space, it's even in my user name. I am glad the barriers of entry are still high but I expect that to change over time as more and more people experiment with sequence learning. But I doubt someone with a few online courses and less than 10 years of trading risk will break into this space. I recommend you find your own niche where you have superior expertise and leave other fields to those who claim domain knowledge that is far superior to yours.
Let's assume 20% of the market participants are profitable. It could be lower than this figure. Are these 20% use neural network? Highly unlikely. So as high frequency. Since the beginning of the Market, is there any system that comes out as a winner consistently? As with any system, it's dependent on the inputs. Garbage in, garbage out. (Thebigshort has already use this statement. Sorry.) Maybe, I should ask this question. If a system that needs good input in order to get good results, is this a good system?
NN is optimization scheme. It works well when data are distributed in space, e.g. flowers, faces, etc. This fails when data is distributed in time for obvious reasons but mainly due to data mining bias. A great article that explains this problem is this one.
thanks to this post i started to research again , maybe i find something anyone know this guy Stefan Jansen - Hands-On Machine Learning for Algorithmic Trading 2018