Univariate Timeseries - which models can be used with ts_learner on prediction on 7 days horizon

[qwertz747]

Hi,

thx again for this great work.
I am having a problem understanding the functionality of ts_learner.
Working with an univariate time series and created the X,y, and split with the following code. So I have a lookbackwindow of 4 weeks and I try to forecast 1 week ahead.

window_length = 7*4 #length of lookback window stride = 1 horizon=7 seq_first=True
X, y = SlidingWindow(window_length, stride=stride, horizon=horizon, seq_first=seq_first, sort_by=average_kg.index, ascending=True)(average_kg)

The x,y,splits have the form:

When I try to build the learner it seems to work only for certain models.
Is there something I am missing, what models can be used for univariate forecasting/prediction regarding a horizon of a week?

For example:
Does not work:

• learn = ts_learner(dls, LSTM, metrics=[mae, rmse], cbs=ShowGraph())
  TypeError: init() got an unexpected keyword argument 'custom_head'

• learn = ts_learner(dls, LSTM_FCN, metrics=[mae, rmse], cbs=ShowGraph())
  TypeError: init() got an unexpected keyword argument 'custom_head'

• learn = ts_learner(dls, MiniRocketPlus, metrics=[mae, rmse], cbs=ShowGraph())
  TypeError: init() got an unexpected keyword argument 'custom_head'

• learn = ts_learner(dls, InceptionTime, metrics=[mae, rmse], cbs=ShowGraph())
  TypeError: init() got an unexpected keyword argument 'custom_head'

Works:
LSTMPlus
GRUPlus

I tried all the same models with TSForecaster too and the same models work or don't work:
batch_tfms = TSStandardize() fcst = TSForecaster(X, y, splits=splits, batch_tfms=batch_tfms, bs=512, arch=GRUPlus, metrics=rmse, cbs=ShowGraph()) fcst.fit_one_cycle(50, 1e-1) fcst.export("models/fcst.pkl")

Can someone please point me in the right direction how to forecast a week ahead with more models (MiniRocketPlus & LSTM_FCN would be great)?

[oguiza]

Hi @qwertz747,
Sorry for my late response.
All models ending in Plus (except MiniRocketPlus which is a special type of model) should work. The Plus actually indicates that additional properties have been added to the original model.

[qwertz747]

Thx for your reply/info!