Neuralforecast different predictions

[whoknowsb]

What happened + What you expected to happen

preds1 and preds2 yield different results. Given an inference_input_size, I should expect the forecast for preds1 and preds2 the same. Am I missing something?

Versions / Dependencies

neuralforecast==1.5.0

Reproduction script

from neuralforecast import NeuralForecast
from neuralforecast.models import LSTM
from neuralforecast.utils import AirPassengersDF

Y_df = AirPassengersDF # Defined in neuralforecast.utils

horizon = 12

models = [
    LSTM(
        h=horizon,
        input_size=48,
        inference_input_size=48,             
        max_steps=500,              
        scaler_type='standard',      
        encoder_hidden_size=64,    
        decoder_hidden_size=64,
        ),   
]
nf = NeuralForecast(models=models, freq='M')
nf.fit(df=Y_df, val_size=64)

preds1 = nf.predict()
preds2 = nf.predict(Y_df.tail(64))

assert((preds1 == preds2).all().LSTM)

Issue Severity

Medium: It is a significant difficulty but I can work around it.

[cchallu]

Hi @whoknowsB! The issue is probably due to the normalization. While the inference_input_size limits the length of the input for the LSTM pass, the normalization statistics are computed on the complete history.

[whoknowsb]

Shouldn't the normalization fitted (mean and std etc...) at training and then applied to inference? Why is calculated different from one to an other? This is really a concerning problem for me. Is normalization calculated on the history dynamically? I think this lead to unespected distribution shifts...