Hola, buenos dias. Vengo en busco de ayuda. Estoy trabajando diferentes tipo de redes neuronales recurrentes para el pronostico de series de tiempo (LSTM-GRU y Simples) , en mi codigo estoy utilizando keras tuner para manejar diferentes hiper-parametros. Ocurre que a pensar de estar utilizando estas herramientas, tengo pronosticos demasiado malos, ya que por ejemplo en esta grafica de casos acumulados, el pronostico es una funciòn decreciente, totalmente fuera del contexto del problema. Espero puedan ayudarme Gracias

Esta es la red neurona que estoy utilizando.

Este es el codigo para mis prediciones de los casos acumulados en 60 dias. Estoy utilizando 670 datos para entrenamiento 60 datos para testeo y ventanas de 60, con batch size de 1. Antes estaba utilizando ventanas de 30 pero los pronosticos eran peoress

Hola @YEISON ARMANDO BUITRAGO LOPEZ. Por lo que veo en tu primera grafica la escala de la Y esta en 1e6, las RNN funcionan mejor escalando los datos, los estas escalando?

Nosotros tenemos ya una RNN implementada en nuestra libreria. En este notebook hay un ejemplo de como usarla: https://github.com/Nixtla/neuralforecast/blob/main/nbs/models_rnn__rnn.ipynb. Podrias probar de usar nuestra implementacion?

Si los estoy escalando, pero luego los regreso a aplico una escalaciòn inversa. Gracias, voy a intentarr

Hello everyone, nice to be here! Big thanks to the team for their contribution(s). Quick question, to be familiar with the package I am planning to use it to model a multi-step multivariate ts problem; any recommendations on which model to try out first? I apologise if this is answered somewhere already.

Hi guys, first of all thanks for your work on this package, looks very promising! I am working with nbeats-like models since the paper came out in 2020 and recently wanted to check out the nhits architecture after its release this year. This was when I came across your package in the first place. While I like the idea of the package, I am struggling a bit with understanding the way you created your Dataset-Classes. From what I understood so far, the two main classes are

TimeSeriesDataset

and

WindowsDataset

. While the first one stores an entire time-series the later one stores sliding windows created from time-series, correct? The second approach feels natural when working with nbeats-like models (as well as other neural forecasting models). However, the implementation you opted confuses me a bit: Previously, when creating my own generators, I have chosen the following recipe for feeding data during training to the model: 1. Sample

n

time-series from the dataset 2. For each series, sample a split time point 3. Create one window for each series form the sampled splits (resulting in

n

windows) and feed them to model as a mini-batch (n = batch_size) This is also the approach described in Oreshkina et al. 2020 (where it is combined with stratified sampling). The implementation you chosen for the

WindowsDataset

(at least as far as I understand it) is different in that all possible windows are generated for a series and retuned (resulting in

n != batch_size

). I am wondering: 1. Why did you choose this implementation style, what are the advantages? 2. Is it possible to implement the Oreshkina et al. 2020 style sampling (especially the stratified version) using your package? Best, Fabian Oreshkina et al. 2020: https://arxiv.org/pdf/2009.11961.pdf

What’s your take on normalization strategies for neural models? Specially when predicting for multiple time series with different scales? I’ve been normalizing per group, but seems that pytorch-forecasting encourages normalizing per observation? Have you experimented with different approaches?

In issue #217

f_cols

is said to be used to identify future available exogenous variables. But I can't find anywhere in the whole code that

f_cols

and

f_idxs

are actually used. How are exogenous variables that are available in the future and the ones that are not, treated differently?

Hello guys. First of all I would like to thank you for this package.

I am trying to use it for some data I got on the web.

It appeared an error

AssertionError: Mismatch in X, Y ds

when performing the following step of "getting_started.ipynb"

Y_df_forecast = model.forecast(Y_df_train, X_df= X_df)

Y_df_forecast.rename(columns={'y': 'y_hat'}, inplace=True)

Y_df_forecast.head()

I checked if Y_df_test had the

ds

correctly following Y_df_train and it has.

Any advice for debugging?

Hi @Andre P, The N-BEATSx model needs you to match the number of observations between

Y_df

and

X_df

, for the moment N-BEATSx only operates with exogenous data available at the time of the prediction. For example: calendar variables, predictions of other series. Can you check if

len(X_df)==len(Y_df)+H

, where H is the forecast horizon (n_time_out)?

Hello @Kin Gtz. Olivares. Yes,

len(X_df)==len(Y_df)+H

.

Hello again. I changed the values of output_size and if they are below 12, I can forecast, otherwise I can't

Any thoughts about it?

is there a model that could forecast multiple time series in this package? perhaps, similar to VAR. thanks

how should we go about making the

S_df

dataframe? i'm on

v0.0.9

and it asks for:

S_df: pd.DataFrame
     Static exogenous variables with columns ['unique_id', 'ds'] and static variables.

i assumed it would look similar to the

Y_df

dataframe, with columns for

unique_id, ds, y

, but on

.fit

it throws the following error:

ValueError('Found duplicated unique_ids in S_df')

i think because we have

unique_id, ds

as a time series in

S_df

, any ideas on what i'm missing?

Hi guys, any clues as to what this error could mean in the context of auto.NHITS?

quick question about AutoNF - does NeuralForecast create new data loaders every time

hyperopt

suggests a new model config? and if so, are each of those loaders guaranteed to contain the same data splits?

is there an easy way to recover val / test predictions from

model_fit_predict

in the same format as

model.forecast

?

Hi @Chris Gervais , if you are using the auto wrappers of the experiment utils: You can parse the predictions from the trials dictionary, see the end part of this N-HiTS Colab example

is there any idea to improve this ?

image.png

the model can't catch up the high peak. I am using gasx model from pyflux. what is the most potential model from neural-forecast to deal with this kind of problem ?

👋 Hello, team!

I am new to neural-forecast and start trying an example: https://github.com/Nixtla/neuralforecast/blob/main/examples/getting_started.ipynb Y_df, X_df, _ = nf.data.datasets.long_horizon.LongHorizon.load('data', 'ILI') while get this error:

TypeError: sort_values() got an unexpected keyword argument 'ignore_index'

any idea what is the issue? Thanks.

Sorry for the inconvenience @Musa Would you be able to open a github issue on your finding and try try this Google Colab meanwhile?: https://colab.research.google.com/drive/1WjBbQzaivQhOldGolzymOtLmo6QX4Ieg#scrollTo=HXKT2-fpUD0Z