Validate is a flat line for values unseen y about time-series-forecasting-tensorflowjs HOT 3 CLOSED

Hi @feuerstarta,

Do check out Why isn't my Model Performing?.

Because the market dipped and recovered within a month or two, this has never happened in history. Since you are using 80/20, the model does not have the chance to learn that pattern.

Also, this is a very simple model, feel free to tweak the model, for better performance. Try building more features on the data as well, possibly adding "difference from previous day".

from time-series-forecasting-tensorflowjs.

See that still makes no sense to be a flat line. the flat line means it is not taking any of the input of the unseen_x into the account. If you do a full predict on the 0-100% of the 80/20 learn, it predicts value till the 80% mark then just flatlines from the last predicted number. If you keep that 100% prediction and do 90/10 split it will go to the 90% mark then flatlines at the 90% mark with the last predicted value. This is showing that the model works only for the section it trains and can not do a prediction period as the model only reports the last recorded train.

from time-series-forecasting-tensorflowjs.

This is the challenge of machine learning, that a model, out of the box, doesn't work for all kinds of data. There is a lot of parameter tuning involve. There may need to do more feature engineering. Or there is a need to change the model architecture.

Doing a quick search, you can see that you are not the only one who faced this problem:

• LSTM always predict same values for all inputs in the batch
• Outputs of LSTM Model all Very Similar or the Same
• Pytorch RNN always gives the same output for multivariate time series

I do not have an answer for you, unfortunately. Some suggestions would be:

• try tweaking the normalize the data portion (.div(tf.scalar(10) might be causing problems line 19 and 20)
• include more data features, maybe can include volume as well
• include more data features, maybe trading indicators such as MACD, RSI, Bollinger bands
• build another model architecture

I'm working on a Python version.

from time-series-forecasting-tensorflowjs.