[core] historic scalers do not properly handle transfer learning

[jmoralez]

What happened + What you expected to happen

When performing transfer learning if we have historic scalers we should refit them to the new data, such that when applying the scaling it has the expected statistics. Currently we reuse the scalers fitted on the original data, which produce different statistics than the expected.

Versions / Dependencies

1.7.2

Reproduction script

import logging
import os

from neuralforecast import NeuralForecast
from neuralforecast.models import LSTM
from utilsforecast.data import generate_series
from utilsforecast.losses import rmse

logging.getLogger('pytorch_lightning').setLevel(logging.ERROR)
os.environ['NIXTLA_ID_AS_COL'] = '1'

# data
series = generate_series(10, min_length=200, max_length=500)
h = 7
valid = series.groupby('unique_id', observed=True).tail(h)
train = series.drop(valid.index)
train2 = train.copy()
train2['y'] += 100
valid2 = valid.copy()
valid2['y'] += 100

# training
nf = NeuralForecast(
    models=[LSTM(input_size=2 * h, h=h, scaler_type=None, max_steps=50, val_check_steps=1, enable_progress_bar=False)],
    freq='D',
    local_scaler_type='standard',
)
nf.fit(train)

# predictions, these should be the same
preds = nf.predict()
preds2 = nf.predict(df=train2)

# comparison
def evaluate(preds, valid):
    return rmse(preds.merge(valid), models=['LSTM'])['LSTM'].mean()

evaluate(preds, valid), evaluate(preds2, valid2)
# (0.7740806216300222, 99.65305363444284)

Issue Severity

High: It blocks me from completing my task.