Two NARX Daily results (during financial crisis time)

Note: window size here decreased to 3. It suited the Daily predicions better. Inputs used are median, close, macd and SMA5.

Advertisements

3 thoughts on “Two NARX Daily results (during financial crisis time)

  1. Having questions regarding your implementation of NARX network for finance:
    1. You’re using delay of 3 for the input and output lines, right?
    2. What normalization routines do you apply to the input and output vectors? ( [-1 1] normalization, maybe some other preprocessing of indicators values? )
    3. Training data:
    input: median(t-1), close(t-1), macd(t-1), SMA5(t-1)
    output: median(t)
    Is this correct understanding of your training process architecture?
    4. What performance measurement function do you use? MSE?

    Thanks,
    Serg.

    Like

  2. Hey Serg,

    I’m trying to remember the exact settings, it’s been quite a while since I’ve played around with NARX networks.

    The input vector consisted of SMA3, SMA5, SMA8 (or SMA5, SMA10, SMA20, you can try both out), in addition to MACD and either median/close. Play around with those combinations, the results barely differ. It’s the most recent data that makes the most difference (SMA3 or SMA5, and PrevClose etc..)

    There was no smoothing or normalization measures. You could try normalizing to [-1,1] or [0,1] and tell me what happens 🙂

    Performance function is MAE, it yielded much better results than MSE. I was thinking of developing my own performance function that is profit/loss-dependent. What do you think?

    Like

    • Hi, Alex,

      Thanks for the reply.
      I’ll test the system according to the information you’ve provided and let you know what I’ve got.
      As of the custom performance function: I’m pretty sure it will boost the performance greatly. I’m going to start research in this field after finding the best combination of input vectors( thinking of some automated script to try out each of the available indicators in Financial Toolbox ).

      Some parameters left unclean for me currently:
      1. Delay for the input and output lines.
      2. Number of hidden neurons and layers.
      Can you advice something on those?

      Regards,
      Serg

      Like

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s