[MLForecast, Lags] Global model. Predicting with lags, individual time series, multiple time series

[mkrech]

What happened + What you expected to happen

Thank you for your fantastic library.

I have a question or bug report about lags and a global model using MLForecast.

My expectation is that in a global model, the results should be identical whether predicting an individual time series or multiple time series, as long as all features are provided correctly and consistently.

However, when I use the lags feature, the forecast for an individual time series (e.g., one time series) differs from predicting the same time series together with multiple other time series.

Versions / Dependencies

mlforecast 0.15.0

Reproduction script

import pandas as pd
import numpy as np
from mlforecast import MLForecast
from sklearn.ensemble import RandomForestRegressor

print("#  Create data")
def create_data():
    np.random.seed(42)
    data = []
    for ts_id in range(3):  # Drei Zeitreihen
        for t in range(30):
            data.append({
                'unique_id': f'ts_{ts_id}',
                'ds': t,
                'y': np.sin(t / 5) + ts_id + np.random.normal(0, 0.1)
            })
    return pd.DataFrame(data)

df = create_data()

print("# single time series")
df_single = df[df['unique_id'] == 'ts_0']

print("# lag features")
lags = [1, 2, 3]

print("# mlforecast model with RandomForestRegressor")
forecast = MLForecast(
    models=RandomForestRegressor(),
    freq=1,  
    lags=lags
)

print("# Prediction single time series")
forecast.fit(df_single, id_col='unique_id', time_col='ds', target_col='y')
single_forecast = forecast.predict(5)  # 5 Zeitschritte vorhersagen

print("# prediction multiple time series")
forecast.fit(df, id_col='unique_id', time_col='ds', target_col='y')
multi_forecast = forecast.predict(5)

print("# filter single time series from multiple time series")
multi_forecast_single = multi_forecast[multi_forecast['unique_id'] == 'ts_0']

print("# results comparison")
print("Single Time Series Forecast:")
print(single_forecast)

print("\nSame Time Series Forecast (with others):")
print(multi_forecast_single)
single_forecast

print("#  difference between forecasts")
difference = single_forecast.set_index('ds')['RandomForestRegressor'] - multi_forecast_single.set_index('ds')['RandomForestRegressor']
print("\nDifference between forecasts:")
print(difference)

Issue Severity

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

[jmoralez]

Hey. You're not "predicting" one or many series, you're training a model with different datasets, so it's expected to produce different results. If you're interested in training one model per serie you can use statsforecast.

[mkrech]

Sorry, you are absolutely right. I can no longer reproduce the issue this way either.
Thank you very much for your support.