The machine operator wrote an article. Basically, я на 86% agree with him
============================
I will continue my vicious attacks on algorithmic trading. Not that, so that I am a supporter of other approaches, and even more intuitive, but the point is this.
In investing / trading, for sustainable results, it is very important not to engage in deliberate nonsense. Nonsense is always an increased risk and, least, waste of time. Cross out the deliberate nonsense from the entire set of activities available to the investor, and you will get more or less working approaches. Unfortunately, most likely, to compile the coveted list, you will have to walk the rake yourself for many years, because there are crowds of loyal fans swarming around any notorious nonsense, charismatic gurus and in general life is in full swing. It's impossible to believe right off the bat, that such a life can boil around nonsense.
So here, как и в прошлый раз, my opinion will be about strategies, which rely solely on price. Now there are a lot of different instruments under a common brand. «машинного обучения». Just grab the libraries you need, stick the price range there, get the model, торгуй, get richer. Want to – neural networks, want to – genetic algorithms. You do not want? well, here's the moving average crossover, not the point. All these Metastocks with Amibrokers – unpretentious multicolored statistical machines «for the humanities».
Important, but, then, that with this approach, an implicit assumption is made about the stationarity of the price series. About that, that the future price series will have parameters, similar to those observed in the past, least, in the near.
And this is where the ambush lurks: row – non-stationary!. And if so, then all this activity of applying super-powerful machine learning algorithms on top of non-stationary data immediately loses its meaning.
I, certainly, far from thought, I'm giving the news. I think, everyone understands everything. Well, or pretend. It's just that the feeling never leaves me, that they carefully close their eyes to this fact, or, at least, keep them not open enough. Deal with their own greed and make a dangerous assumption.
And the fact deserves it, to be looked at as closely as possible, because the nonstationarity of the series gives a very bad effect for risk management:
You cannot assess future risks.
You can not, because you just don't know the parameters of the future series. Потому что он, infection, non-stationary. I'm not even talking about future profitability. Profitability is ultimately not that interesting, how risks. If only because, that in the long term risks tend to strongly affect profitability.
If algorithmic trading is no more for you, what an exciting hobby, it's not scary. Well, it will bomb your equity, maybe you came here for adrenaline. But if this is a profession, this kind of risk uncertainty may reflect badly on your statement. You trade, for example, ten years old with a beautiful sharp, and then some strategy goes crazy and the end of your sharp, and possibly, карьере.
What to do?
My version – it is necessary to look for contexts, where is a priori(!) return to mean. Парный Trading похожих активов, for example. Flows between markets or sectors. Risk-on, risk-off dynamics. Что-нибудь такое, where the stationarity of the series is provided by the very logic of the process.
