Hey everyone! Thank you for all the amazing work and for continuing to keep it updated. I saw that recently a PR was approved for predict_insample. Does anyone have any examples on how to use it? I'm attempting to use it but keep getting an error. I'd love to see some examples if anyone has used it successfully out there!

Hey everyone! I often receive the following error while training AutoNHITS

IntegerGreaterThan(lower_bound=0)

with the following config: What are the reasosns?

horizon = 60

config_nhits = {
    "input_size": tune.choice([14 ,28, 28*2, 28*3, 28*5, 2 * horizon]),              # Length of input window
    "n_blocks": 5*[1],                                              # Length of input window
    "mlp_units": 5 * [[512, 512]],                                  # Length of input window
    "interpolation_mode": tune.choice(['linear']), 
    "n_pool_kernel_size": tune.choice([5*[1], 5*[2], 5*[4],         
                                      [8, 4, 2, 1, 1], [16, 8, 1]]),            # MaxPooling Kernel size
    "n_freq_downsample": tune.choice([[8, 4, 2, 1, 1],
                                      [1, 1, 1, 1, 1],
                                      [168, 24, 1], 
                                      [24, 12, 1], 
                                      [1, 1, 1]]),            # Interpolation expressivity ratios
    "learning_rate": tune.loguniform(1e-4, 1e-2),                   # Initial Learning rate
    "scaler_type": tune.choice([None]),                             # Scaler type
    "max_steps": tune.choice([1000]),                               # Max number of training iterations
    "batch_size": tune.choice([16, 32, 64, 128, 256, 512]),                  # Number of series in batch
    "windows_batch_size": tune.choice([32, 64, 128, 256, 512, 1024, 2048]),       # Number of windows in batch
    "random_seed": tune.randint(1, 20), 
    "scaler_type": tune.choice(["robust", None]),                            # Random seed
    "hist_exog_list": ["week_day", "month", "trends"],
    "futr_exog_list": ["week_day", "month"]
}

I am trying to use Temporal Fusion Transformer (TFT) (https://nixtla.github.io/neuralforecast/models.tft.html#tft). Can I extract variable importances from NeuralForecast's implementation of TFT? I am referring to Pytorch's implementation of TFT (https://pytorch-forecasting.readthedocs.io/en/stable/tutorials/stallion.html). Thanks.

On the topic of TFT, any suggestions on speeding up the training (besides GPUs obviously)? There seems to be some bottlenecks, maybe on the dataset loader side?

Hello! I'm just getting started with the library and having a lot of fun with it! My goal is to set up a model that predicts 24 hours worth of prices every day. I was originally going to do this by training up to some date with horizon=24, and for the next N days, call

nf.predict(futr_df=<day-to-pred>)

However, it seems that this always returns a dataframe with a ds column with just the next 24 hours after where I stopped training. Is there some way to apply to model to the next N days without retraining it every time? or would i have to retrain / finetune on the data since then? Perhaps the latter is the only way to support historical features?

is there any constrain on the period i should predict on the future?

Can you explain more about max_steps parameter in NHITS? As I increase the no. of max_steps (from 1 to 100), the model gets more accurate, however, I am unable to replicate my predictions with higher no. of max_steps.

Hi! I'm currently working my way into NeuralForecast and have a somewhat basic question I couldn't find any documentation about: When I'm creating my timeseries, let's say daily sells of a certain product, should I add a y=0 for dates the product hasn't been sold at all or can I just skip the date in the time series? I know it works with gaps in the time series, but I'm wondering if the result is accurate or if that distorts the model

Having trouble thinking through how to fit a current forecasting problem into the neuralforecast model: Suppose every day at 10am, I want to predict tomorrows 24 hours worth of prices for some series (midnight to midnight). It seems that currently the assumption is if I'm predicting time t -> t+horizon, I have data up to t-1. But in this case, I only have data up to

t-14

as i do not yet have prices from 11am-midnight I could extend the horizon to be

24+14

to predict 10am -> 12am the next day, but then in training / fit we'd step by the horizon, instead of just stepping forward 10 days. is there some way to fit this into the neuralforecast approach?

Hey, maybe you can help me here: I use Auto-NHITS with the standard parameters to forecast the following timeseries: Unfortunately, I get an MAPE of 35%. With which parameters should I play to increase the accuracy?

Hi! I have a quick question about NBEATSx. I noticed that the both exogenous basis detailed in the paper aren't currently available in neuralforecast atm so I'm trying to add them in myself based on the research paper repo. For these two basis, what is the "proper" way to use historical exogenous variables? It looks like the research code handles all variables as future exog variables (unless I'm mistaken). Thanks again for the help!

Looking at adding the HINT architecture tomorrow, will circle back with metrics when available. Any thoughts on a reasonable hparam search space? Didn’t see anything added to the auto module

Hi, I just want to make sure I fully understand everything. I have the following config dictionary

mlp_config = {
        "input_size": tune.choice([2, 4, 12, 20]),
        "hidden_size": tune.choice([256, 512, 1024]),
        "num_layers": tune.randint(2, 6),
        "learning_rate": tune.loguniform(1e-4, 1e-1),
        "batch_size": tune.choice([32, 64, 128, 256]),
        "windows_batch_size": tune.choice([128, 256, 512, 1024]),
        "random_seed": tune.randint(1, 20),
        "hist_exog_list": tune.choice([pcc_list]),
        "futr_exog_list": tune.choice([fcc_list]),
        "max_steps": tune.choice([500, 1000]),
        "scaler_type": tune.choice(["robust"]),
    }

The

hist_exog_list

and

futr_exog_list

are lists of historical and future exogenous variables. When I fit the model I use

nf.fit(df=df, val_size=20)

where the

df

is my dataframe that contains the target variable as well as the exogenous variables. Then when I predict I use

nf.predict(futr_df=futr_df)

where

futr_df

contains only the observations of the future exogenous variables that extend beyond the time period of the point of prediction. Does this seem correct? Or am I doing something wrong in the specification of the future exogenous variables. In my case there is only one period beyond the cut-off for the target where future exogenous variables are available, so the

future_df

dataframe only has one row and many columns (for the different features).

hello i get this error when try to use cross validation with automodels RuntimeError: maximum size for tensor at dimension 2 is 2751 but size is 3360 does any one have idea?

Hi, I was wondering if I had enough data to try any model in neural-forecast. I have 176 months history and I need to forecast 24 months into the future. Is it enough? Or not suitable? Which model would you recommend to test? Thank you very much.

Hello I have a question with regards to how the unique_id impacts NeuralForecast. I have been using the NHITS model with hierarchical data like in the first image and when I use the aggregate function from the HierarchicalForecast to create the data frame the data looks like the second image. I noticed when using the unique_ids from the HF aggregate function I get different forecast results despite everything else being 100% the same. This leads to RMSE values that are substantially different. The current work around I have is to just rename the unique_ids back to the aggregated ones given by the aggregate function from HF once I get the forecasts but just wondering if anyone might have any insight as to why this is happening and if its supposed to be behaving this way. Thank you!

Is their a way to do transfer learning using neural forecast library?

Hi all, I am relatively new to time series forecasting, and wondering what is the proper way to treat to multiple independent time series. Consider the example of k data series x_{t_0}^{t_i}[k] from time t_0 to t_i (e.g. stocks price for different stocks, or data series data for different agents), and we want to predict their results for the next T steps x_{t_i+1}^{t_i+T+1}[k]. We cannot simply append these series as in text data, since we want to preserve the time information across series. I think a naive way to use the unique_id as the identifier for the different series. However, unique_id is a dataset feature, and I am not sure how it is treated for different models. And if it is always the suggested way to handle univariate multiple (independent/dependent) series?

Is their a way to add embedding vector or glove vector to neural or ml models in nixtla?

Hi all, Is there any notebook example in which I can see how to implement AutoNBEATSx? So far I have seen this (for NBEATSx without hyperparmenter tunning) and this (for hyperparameter tunning without exogenous variables). For example, when I run this: models = [AutoNBEATSx(h = n_day_forecast, loss=MAE(), config=nbeats_config, input_size = 5*24, futr_exog_list = fut_exog_list, hist_exog_list = hist_exog_list, stat_exog_list = static_list, scaler_type = 'robust', search_alg=HyperOptSearch(), num_samples=20)] I have the following errors: TypeError: AutoNBEATSx.__init__() got an unexpected keyword argument 'input_size' TypeError: AutoNBEATSx.__init__() got an unexpected keyword argument 'futr_exog_list' Thank you very much!

is there function built in neuralforcat to check seasonality

When loading from checkpoint why getting this error NotImplementedError ("{} cannot be pickled", MultiProcessingDataLoader)

I don’t understand why my chart Y-axis is like that, any hints?

3,1,2 makes no sense in my eyes 😄

When running nf.predict() for vanilla transformer, I got "can only concatenate str (not "pandas._libs.tslibs.offsets.Hour") to str", any hints? I am following xxxx-xx-xx xxxxxx format. The fit function works fine though. I think there is some issue with the format.

Good day, I seem to be getting an error

RuntimeError: einsum(): subscript p has size 2 for operand 1 which does not broadcast with previously seen size 0

. This only happens with the AutoNBEATS and AutoNBEATSx models. Other models run fine. The code that I have for AutoNBEATS worked perfectly fine until about a week ago.

Can a global model trained on data with a freq of 1 hour be used to predict a similar data but with a freq of 10 mints?

Hey everyone, I’m considering to train a pre-trained model for our use case (5 years of media data). I was wondering: is there more information about the training set for the pretrained models on https://github.com/Nixtla/transfer-learning-time-series?

Hi, I am wondering what is the proper way to perform in-sample evaluation? For example, the transformer-based model is trained to learn within window behavior, where we are given data

input_size

, but asked to make prediction on a data of

h

. The goal is to make evaluate over randomly chosen window on the validation dataset. Currently, I am thinking of doing resampling myself, but I am wondering if there is in-built features for insample evaluation. I think

predict_insample

might be doing this, but I am confused about what the output of predict_insample is. Does it only contain the predicted value? Is it a fair metrics to simply compute accuracy of predict_insample with the original df as a measurement of method.

hello! first, thanks for putting together a great library and running this community, i really appreciate it! i had a question about one of the NHITS hyperparameters,

n_freq_downsample

. i noticed in the AutoNHITS default_config definition, there is a tune.choice over the following values:

"n_freq_downsample": tune.choice(
            [
                [168, 24, 1],
                [24, 12, 1],
                [180, 60, 1],
                [60, 8, 1],
                [40, 20, 1],
                [1, 1, 1],
            ]
        ),

do you all have any intuition about whether lining up these frequencies with known natural frequencies of the data is useful for performance? for example, [168, 24, 1] seems to correspond to weekly (24 x 7) , daily (24 x 1), and hourly frequencies. the reason i ask is as follows: let's say i have an NHITS model that predicts hourly-sampled data, and i find through the course of hyperparameter optimization that

n_freq_downsample=[168, 24, 1]

is most performant. if i was then to train a model that predicts the same series, but now sampled at 10minute frequency (6 samples per hour), should i then change my hyperparameter search space to include a choice for

n_freq_downsample = [168*6, 24*6, 6]

? any insight you might have would be appreciated - thanks in advance!