Are multi-step predictions possible?

kaelzhang / DA-RNN-in-Tensorflow-2-and-PyTorch

A Tensorflow 2 (Keras) implementation of DA-RNN (A Dual-Stage Attention-Based Recurrent Neural Network for Time Series Prediction, arXiv:1704.02971)

MIT License

30 stars 12 forks source link

Are multi-step predictions possible? #2

Open moritz-raabe opened 3 years ago

moritz-raabe commented 3 years ago

Hey,

thanks for providing this nice library! :)

Running your example I was wondering if it is possible to change the length of the predictions. As far as I have seen, only a prediction of length 1 is produced. But for my thesis I would like to forecast up to 96 time-steps into the future...

Would that be possible and if yes could you provide an example?

Thanks in advance and have a great day!

kaelzhang commented 3 years ago

Generally, there are two kinds of multi-step predictions:

single-shot model: it is controlled by y_dim. For example, if you want to do 2-step no-shift predictions, you are required to preprocess training data with y_dim=2
autoregressive model: it is not mentioned in the paper, and da_rnn does not support regression and it only could predict y_T_hat, see here.

kaelzhang commented 3 years ago

Besides, IMO, time-series prediction is a very bad way to do quant trading and multi-step predictions will bring even worse results.

I think a 96 time-step prediction will be worse than doing nothing.