[Core]ts.update method may not support target_transforms,lag_transforms and date_features

[MrTangsai]

What happened + What you expected to happen

Hi, me again, long time no see.
I want to maintain a stable model and update subsequent prediction results based on new data every day, I use ts.update() to get it , but encounter the following issues:

1. When transformed the target like target_transforms=[Differences([1])] then update the subsequent value, but the data used in the actual model has not been differentiated, I saw the source code, under update method, self.ga changed but not consider the situation when if self.target_transforms is not None
2. When add lag_transforms or date_features , get a bug TypeError: 'Series' object is not callable

Versions / Dependencies

1. mlforecast: 0.11.4
2. python: 3.8.16
3. os: Ubuntu

Reproduction script

situation 1 code is below:

import pandas as pd
from mlforecast import MLForecast
from mlforecast.target_transforms import Differences

from lightgbm import LGBMRegressor
from IPython.display import display
def inspect_input(new_x):
    """Displays the model inputs to inspect them"""
    display(new_x)
    return new_x
df = pd.read_csv('https://datasets-nixtla.s3.amazonaws.com/air-passengers.csv', parse_dates=['ds'])
fcst = MLForecast(
    models=[LGBMRegressor(verbose=-1)],
    freq='MS',  # our serie has a monthly frequency
    lags=[1, 12],
    target_transforms=[Differences([1])],
)
fcst.fit(df, )
print(fcst.predict(1, before_predict_callback=inspect_input))
new_df = pd.DataFrame([['AirPassengers', '1961-01-01', 500]], columns=['unique_id', 'ds', 'y'])
new_df.ds = pd.to_datetime(new_df.ds)
fcst.ts.update(new_df)
print(fcst.predict(1, before_predict_callback=inspect_input))

then the result is below:

situation 2 code is below:

import pandas as pd
from mlforecast import MLForecast
from mlforecast.target_transforms import Differences

from lightgbm import LGBMRegressor
from window_ops.rolling import rolling_mean, rolling_std

from IPython.display import display
def inspect_input(new_x):
    """Displays the model inputs to inspect them"""
    display(new_x)
    return new_x
df = pd.read_csv('https://datasets-nixtla.s3.amazonaws.com/air-passengers.csv', parse_dates=['ds'])

fcst = MLForecast(
    models=[LGBMRegressor(verbose=-1)],
    freq='B',  # our serie has a monthly frequency
    lags=[1, 12],
    # lag_transforms={
    #     1: [(rolling_mean, 5), (rolling_mean, 3), (rolling_std, 5), (rolling_std, 3)],
    # },
    date_features=['dayofweek'],
    target_transforms=[Differences([1])],
)
fcst.fit(df, )
print(fcst.predict(1, before_predict_callback=inspect_input))
new_df = pd.DataFrame([['AirPassengers', '1961-01-01', 500]], columns=['unique_id', 'ds', 'y'])
new_df.ds = pd.to_datetime(new_df.ds)
fcst.ts.update(new_df)
print(fcst.predict(1, before_predict_callback=inspect_input))

then the result is below:

Issue Severity

High: It blocks me from completing my task.

[jmoralez]

Hey @MrTangsai, thanks again for the excellent report and sorry for the troubles. #293 should fix both errors. We'll make a release soon and in the meantime you can install from github (pip install git+https://github.com/Nixtla/mlforecast).