MLforecast predicting on unseen unique ids & predicting on future seen ids.

[github-abdellah-elyounsi]

Hi there, thank you for this amazing lib.

I have 2 questions please:

1. I'm trying to figure out how to use MLforecast for new unseen unique ids. My understanding global models should be able to predict new time series that have not been in the training data using static and dynamic features, but i'm having hard time figuring out how this is supposed to work.

2. if ihave a time in training data say starts from 2020-01-01 till 2020-02-01 and use dynamic features. My test data starts in 2020-03-01 as I dont have the data for the gap between 02/01 and 03/01. How can I do predict on the test set ?

Here is what I tried for first question:

import pandas as pd
import numpy as np
from mlforecast import MLForecast
from catboost import CatBoostRegressor

np.random.seed(42)

dates = pd.date_range("2025-01-01", periods=10)
ids = ["A", "B"]  # training IDs

data = []
for uid in ids:
    for d in dates:
        data.append({
            "category": uid,
            "unique_id": uid,
            "ds": d,
            "y": np.random.rand() * 10
        })

df_train = pd.DataFrame(data)

lags = [1, 2]

model = CatBoostRegressor(
    iterations=100,
    learning_rate=0.1,
    depth=3,
    verbose=0,
    cat_features=['category'],
    random_state=20,
)

fcst = MLForecast(
    freq="D",
    models=model,
    lags=lags,
    date_features=["dayofweek"]
)

fcst.fit(df_train, id_col="unique_id", time_col="ds", target_col="y", dropna=False)


-- forecast a new unseen timeseries.

new_id = "C" 
future_dates = pd.date_range("2025-01-11", periods=5)

new_df = pd.DataFrame({
    "unique_id": new_id,
    "ds": future_dates,
    "category": new_id,
})

preds = fcst.predict(h=2, new_df=df_new)

---> ValueError: The following columns are missing: ['y']

This is kinda suprising because why would you need the target for to forecast ? to compute the lags ?

I added the column to my new_df but and run .predict()

The suprising thing is that it trained on that data and forecasted day 16 and 17.

For the second question, if you are predicting using a gap I get the following error.

ValueError: Found missing inputs in X_df. It should have one row per id and time for the complete forecasting horizon.
You can get the expected structure by running `MLForecast.make_future_dataframe(h)` or get the missing combinatins in your current `X_df` by running `MLForecast.get_missing_future(h, X_df)`.

[github-abdellah-elyounsi]

Could you please help ?