Hello, I am trying to predict values of unseen data based on the historical data of a similar products. Example: I'm trying to predict the sales of a new shirt that will be released in summer, I already have data from the sales of a similar type of shirts in previous year(s)/month(s). Based on the guidance by Kin it appears that this can be achieved using Temporal Fusion Transformer (https://nixtla.mintlify.app/neuralforecast/models.tft.html), however, I am struggling in its application. • My assumption is that

unique_id

does not play any role in the following model and can be simply set as

"shirts"

, with overlapping/repeating dates (

ds

) from previous shirt sales? NOTE: There are exogenous variables available (for the sake of an example), such as weather, humidity and number of people in the city. Example application using AirPassenger dataset:

import pandas as pd
import pytorch_lightning as pl
import matplotlib.pyplot as plt

from neuralforecast import NeuralForecast
#from neuralforecast.models import TFT
from neuralforecast.losses.pytorch import MQLoss, DistributionLoss, GMM, PMM
from neuralforecast.tsdataset import TimeSeriesDataset
from neuralforecast.utils import AirPassengers, AirPassengersPanel, AirPassengersStatic

#AirPassengersPanel['y'] = AirPassengersPanel['y'] + 10
Y_train_df = AirPassengersPanel[AirPassengersPanel.ds<AirPassengersPanel['ds'].values[-12]] # 132 train
Y_test_df = AirPassengersPanel[AirPassengersPanel.ds>=AirPassengersPanel['ds'].values[-12]].reset_index(drop=True)

Modification:

Y_train_df["unique_id"] = "Airline1"
Y_test_df = Y_test_df[lambda x: x["unique_id"] == "Airline1"]

Modeling:

nf = NeuralForecast(
    models=[TFT(h=12, input_size=48,
                hidden_size=20,
                #loss=DistributionLoss(distribution='Poisson', level=[80, 90]),
                #loss=DistributionLoss(distribution='Normal', level=[80, 90]),
                loss=DistributionLoss(distribution='StudentT', level=[80, 90]),
                learning_rate=0.005,
                stat_exog_list=['airline1'],
                #futr_exog_list=['y_[lag12]'],
                hist_exog_list=['trend'],
                max_steps=500,
                val_check_steps=10,
                early_stop_patience_steps=10,
                scaler_type='robust',
                windows_batch_size=None,
                enable_progress_bar=True),
    ],
    freq='M'
)
nf.fit(df=Y_train_df, static_df=AirPassengersStatic, val_size=12)
Y_hat_df = nf.predict(futr_df=Y_test_df)

Hi @Akmal Soliev 🙂. Yes, it should be possible to use the sales of a different

unique_id

to predict a new article (in fact, any model would work). However, the change should only be done during the

predict

step. My recommendations are: 1. Fit your model on the historical values of your existing shirts, as you would normally do. There is no need to change any

unique_id

. Do not change and group the different products in the same

unique_id

. Each

unique_id

should only have one value for each

ds

. 2. When using the

predict

method, pass a new

df

using the appropriate values for

y

from a different article. For instance, set the dates for the current year but take the values for

y

of the existing similar shirts from the previous year. (Of course, this approach is making many assumptions to work properly). Let me know if this is clear.

Yes, it should be possible to use the sales of a different

unique_id

to predict a new article

In this case if I have passed

blue shirt

red shirt

white shirt

as train unique id I'd still be able to predict

yellow shirt

? I think if there was a minimal example that would help a ton! Thanks Christian, struggling to wrap my head around this idea.

With neuralforecast models you can use a trained model to forecast any time series you want by simply passing a new

df

in the

predict

method. The

unique_id

of the

df

does not have to be contained in the train data. By training on

blue shirt

red shirt

white shirt

, etc. the model has learned to forecast multiple types of shirts. You can then construct your

df

for the predict method with a new

unique_id

, for example

yellow_shirt

, and take the

y

values of a "very similar" shirt. Or maybe produce multiple forecasts for several "very similar" shirts and then take the average/mean.

1. When using the

predict

method, pass a new

df

using the appropriate values for

y

from a different article. For instance, set the dates for the current year but take the values for

y

of the existing similar shirts from the previous year. (Of course, this approach is making many assumptions to work properly).

okay i literally just had to change the parameter 🤦, was so confused on why you said

df

😄 Does it work with exo variables? I attempted to run this on Airpassanger example from above with the following variation to

Y_test_df

:

unique_id	    ds	        y	  trend	y_[lag12]
0	TestingAirlines	1960-01-31	417.0	132	360.0
1	TestingAirlines	1960-02-29	391.0	133	342.0
2	TestingAirlines	1960-03-31	419.0	134	406.0
3	TestingAirlines	1960-04-30	461.0	135	396.0
4	TestingAirlines	1960-05-31	472.0	136	420.0

I am getting the following error:

Exception: {'airline1'} static exogenous variables not found in input dataset

Yes, it allows for exogenous. In the dataframe for the predict method you must include all variables added when taining

all features? If yes, then all features are identical. The only difference is the

unique_id

Yes, if you pass all the same features then the forecasts will be the same. You essentially assume, for example, the forecasts of the

yellow_shirt

will be identical to the

blue_shirt

. However, depending on your exogenous features you might have different behaviour (such as weather or number of people for the new year). Another option as I mentioned is to take some form of aggregation from multiple forecasts for different similar products (it has a k-nearest neighbor flavor).

I have attempted the first option, but it gives me an error :/ Second option is nice, will look into it after the first one

One last option is to leverage TFT's static_features in the encoder directly. You should first create several static features to characterize products (for example category, color, size, etc.). Then use static features + calendar variables to produce forecasts without past history.