Weird drop in predictions

[ricokoff]

Hi,

I am trying to get your model to work with financial data for my Thesis. However, the predictions seem to be shifted by quite a bit. I added an image to show this. Do you have any idea what could cause this shift? I am currently doing multivariate predict univariate (MS) with 6 features, including the target. prediction.pdf

Some of my parameters are: --features MS --target close --freq h --seq_len 96 --label_len 48 --pred_len 96 --enc_in 6 --dec_in 6 --c_out 1 --d_model 512 --n_heads 8 --e_layers 2 --d_layers 1 --d_ff 2048 --dropout 0.05 --embed timeF --activation gelu --num_workers 10 --itr 2 --train_epochs 10 --batch_size 32 --patience 3 --learning_rate 0.0001 --loss mse

Thanks for your efforts and I hope you can help me out!

[zhouhaoyi]

You can provide a sample of your dataset. We may help checking.

[ricokoff]

BTC_USD_CEX.csv Great, here is my current dataset

[ricokoff]

@zhouhaoyi Could you find the cause of this problem? I tried several different (financial) datasets, but I keep getting the same problem. Maybe I am using the wrong parameters or I need to preprocess the data more? Please let me me know if you found something.

[zhouhaoyi]

We are checking on the data. Please watch our repo for updates.

[ricokoff]

@zhouhaoyi any updates? I need to get this working for my thesis.

[ricokoff]

BTW, in the prediction.pdf file I provided, you can see what I mean with the weird drop

[MountVoom]

The distribution of your training and testing sets of data is very different, and it will be difficult to predict.