[patryknextdoor]

As you agreed with me that pattern does not mean the well known chart patterns. Patterns means time series patterns that have some statistical property that can be measured.

Can you teach me how to extract and interpret useful information/patterns from time series with some statistical measurable property for trading?

[patryknextdoor]

Good Evening Sir,

Can i still hope for your reply regarding this thread post? I will highly appreciate your help on that topic.Thank You in advance.

Best regards,

P.

[patryknextdoor]

Can i still hope for your reply regarding this thread post?

[Hassam]

ARIMA is a popular method used in statistics books for analyzing the trend in a time series. In ARIMA process requires first making the time series stationary by differencing. Differencing is done so that the time series mean becomes zero and we have a stationary time series with a mean that does not change with time. When you do differencing which is simply subtracting the present value with the previous value, you lose the time series memory. So this is what I have found ARIMA does not make good financial time series forecasting. I have also tried to use Deep Neural Networks but doesn't work either. So I don't waste time on using ARIMA or DNN on time series data. Every moment something is happening and the time series is changing its behavior so it is difficult to make good predictions using these models. Maybe there are people who have built good time series models. As far as I am concerned my time series models did not work.

So why can't we make good time series models. This is what I think why we can't make good time series models. We only have price data. We don't know the exogenous variables that are impacting that price. Breaking news can be one. Once we don't know the other variables that are impacting price, it becomes difficult to build good time series models.

[patryknextdoor]

1.Reading the book by Graham Giller (Adventures in Financial Data Science) he mentioned about state space models the importance of choosing some non-stationary models. How about that?

2.Whats a good practice to at least try to find and define the key variables that might move price at the current day? Is there some practices for that?

3.Are you still going to create the algoritmic trading course with python from A to Z ?
and signal chat group?

i would be very happy and grateful to learn from you.

please let me know

best regards

[Hassam]

I have build state space models as well. The problem with state space models is that we use normal distribution as usual which  does not work in practice. We also need to choose the type of state space model for example we need to specify whether the model is linear first order, linear second order, linear third order. State space models are bayesian in nature meaning each new observation is incorporated sequentially. Particle filters were supposed to be solution. But I stopped using state space models as I was not getting good results. Think like a sharp shooter. What model the sniper uses. He just uses practice to pull the trigger.

I plan to start the python algorithmic trading course on my youtube channel soon. I will also build the signals app on youtube.

[patryknextdoor]

Thank you very much. I cant wait already for the course! Very excited to finally learn from expert.

As Graham Giller mentioned in his book, that he try to avoid bayesian stats and he is using FISHERIAN (frequentist) way of statistics. maybe thats a way to go about it since the assumptions of Fisher stats are different from Bayesian.

[Hassam]

I am a firm believer in Bayesian statistics. It makes more sense. You start with a prior belief. With new data you update your belief. This makes more sense.