Discussing the article: "Self-Learning Expert Advisor with a Neural Network Based on a Markov State-Transition Matrix" - page 2

 

Interesting approach, but the text is pretty weird to me at some places, it is extremely overoptimistic, and  some wordings are clearly AI-written ( "This code is a genuine alchemical laboratory where raw market data is transformed into a precious elixir of knowledge" — what even is this sentence?). Also I have some severe structural and mathematical problems with this approach. I have been working with Markov-matrices before, and let me tell you, 0.67 as diagonal value (persistence) is clearly not a stable state! 0.67 means the asset averages only 3 days in the same regime before flipping, which is completely unstable and just captures short-term technical noise, especially since you are using these regime-identifications on assets which are typically not too trending.

Also I don't understand how using a rolling window of 100 bars can help the Markov-matrix react quickly to regime shifts. On a daily timeframe, a rolling window carries over 99% of the exact same data from one day to the next. The resulting transition matrix is highly autocorrelated and moves like a sluggish, heavily lagged moving average, meaning it will smooth out and lag behind a structural market break rather than reacting instantly.

Also speaking from experience in case of Markov-models, using more than 3 states (even 4-5 is mostly too much) just creates degenerated models through hat-singular solutions, where a state does not even have any data density assigned to it, or a transition is not even recorded, so the model gets too complicated to represent anything meaningful.

Feeding a 9-element flattened matrix of slow-moving historical frequencies into a 40-hidden-neuron Multi-Layer Perceptron to predict immediate, next-bar price direction seems highly counterintuitive. The input features are highly static day-to-day, forcing the MLP to overfit on the 600 training samples. So I dont know about this whole approach using MLP based on Markov-matrix, it seems a bit odd, and the backtest even odder based off of it, but good luck on making it work!