Strange results about time-series-forecasting-tensorflowjs HOT 8 CLOSED

Or may someone can recommend any content to study and understand it better? Thanks

from time-series-forecasting-tensorflowjs.

Hi, could you upload a more complete picture of the graph, with the legends and the axis. And also the parameters you have used.

from time-series-forecasting-tensorflowjs.

Hi, could you upload a more complete picture of the graph, with the legends and the axis. And also the parameters you have used.

Hello! In attach the JSON containing EURUSD (~6.000 5 minutes candles) data:
https://drive.google.com/file/d/1XXiUfSL9eBkPjUhAnu4mTkH1jdceieTz/view?usp=sharing

I am using the same algorithm with the following parameters:

Training Dataset Size: 98
Epochs: 25
Learning Rate: 0,01
Hidden LSTM Layers: 4
Simple Moving Average Period: 20

I also tried with different training dataset size, epochs and SMA periods... The result are always the same.

This is the training chart:

This is the validation chart:

And this is the predict chart:

This trained model download:
https://drive.google.com/file/d/1n7eMn-OTMK1ziSqI-2IOKOFR3jLrm_ZH/view?usp=sharing

from time-series-forecasting-tensorflowjs.

I am facing exactly the same problem

from time-series-forecasting-tensorflowjs.

I am seeing the same result as well, i forked a version and trained the code, i see the same result locally as well.
The only difference between your initial commit is this 2 lines.

let X = inputs.slice(0, Math.floor(trainingsize / 100 * inputs.length));
let Y = outputs.slice(0, Math.floor(trainingsize / 100 * outputs.length));

The current version doesnt have that, not sure if its anything to do with the issue. (I am still learning, so not sure if thats the issue).

Also on this link https://jinglescode.github.io/time-series-forecasting-tensorflowjs/
Followed all the settings you used on your readme, and i also got a flatline for validating the prediction.

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.

Indeed, 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
• PyTorch regression is producing the same numbers as prediction

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

from time-series-forecasting-tensorflowjs.

@kaykyr, it's really strange you are getting a flat line on the training set too. I think have to debug (doing lots of console.log) and figure out what is the issue. At the minimum, you should get the model to overfit on the training set.

from time-series-forecasting-tensorflowjs.

@kaykyr, it's really strange you are getting a flat line on the training set too. I think have to debug (doing lots of console.log) and figure out what is the issue. At the minimum, you should get the model to overfit on the training set.

Hey, thank you!

It's very strange, but I built the same model with python with latest versions (Keras/Tensorflow) I got other result, but looks like it's predicting to the future the same thing of the past.

But, it's ok, now I know that this is a thing that needs too much effort to works great, I'll keep learning and improving my model.

from time-series-forecasting-tensorflowjs.