I have a set of time series data which gives me fibonacci levels and the duration at which the value is at this level. Data structure to look like:

Date / Duration (minutes) / Level

201201 / 380 / 2

210422 / 400 / 4

I'd like to create a NN model (LSTM maybe) that would forecast the next level, the probability it reaches it and this for several steps ahead (1 step = 400 minutes). Which time series forecasting model would you recommend ? Thanks in advance.

  • $\begingroup$ Could you explain what a Fibonacci level is? I'm not sure what you are trying to achieve with your model. What is the desired output? $\endgroup$
    – Marcus
    Apr 28 at 19:34
  • 1
    $\begingroup$ Fibonacci levels in stock price (for example) is defined between the minimum and the maximum price during a given period. The range (min, max) is chunck in levels. Below min and above max levels are called extension levels. So, let's say that I look at 1 year back. It gives me my level and I start classifying and I count the duration in minutes price stays in each level. I am now trying to have a time series prediction model which tells me the probability that in X minutes, it should be at the level Y with a given probability. Hope it's clear, if not I am happy to explain more. $\endgroup$ Apr 28 at 19:38
  • $\begingroup$ So basically, instead for predicting the actual value of the stock price, you want to know in which range the price will fall in and for how long? $\endgroup$
    – Marcus
    Apr 28 at 19:44
  • 1
    $\begingroup$ Exactly this with a given probability $\endgroup$ Apr 28 at 19:45

It seems like you are trying to produce two outputs here. This usually makes the models more complex. What if you only predicted the Fibonacci levels for each time step? Then count how many time steps it stays at that level.

As for producing the Fibonacci levels themselves, you can look into categorical time series where the values are categories instead of real numbers.

You have a many-to-one problem. So, yes the model takes two inputs, Duration and Level, but only produces one output, the Level.

The times septs in a model will always be the same. You have to choose a range that is appropriate and all your predictions will be based on that. It's important to note that finding the best time step size is a trial and error process. For example, you could first try using 10 time steps and if you find the model needs more you retrain it with 20 and compare the results.

  • 1
    $\begingroup$ So, I have basically two input for each time step (duration, level) and the output is the next level only ? 2nd question, does the time step have to be equal ? Thanks for your help $\endgroup$ Apr 28 at 20:07
  • 1
    $\begingroup$ Thanks a lot. Could you please clarify one thing more. Do you confirm that a step is a duration regadless of the duration: i.e input vector (300 minutes, Level 2) output (Level 3) and 1 step is this 300 minutes, but could be 222 or anything else ? or one step is a given duration ? Do you recommend Rocket as time series forecasting for this exercise ? $\endgroup$ Apr 30 at 12:43
  • $\begingroup$ @user3311337 Yes you are correct about the steps. However, in time series a time step is how far back you are considering. For example, using one week to forecast the next day is a time step of either 7 days or 1 week. Depends on your data. As for Rocket, I never used it so I can't recommend it. $\endgroup$
    – Marcus
    Apr 30 at 13:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.