predict_in_sample with exogenous variables

[bloukanov]

Describe the bug

If an ARIMA model is fit with exogenous variables, in-sample predictions do not appear to depend on the X values provided to predict_in_sample. In fact, even arrays with missing columns and rows are allowed. Are the true X values in sample saved somewhere, so they don't need to be provided? Or am I missing something here..?

To Reproduce

import pmdarima as pm
from pmdarima import model_selection
import numpy as np
import pandas as pd

np.random.seed(42)

y = pm.datasets.load_wineind()
df = pd.DataFrame(
    {
        "x1": y * np.random.uniform(0, 0.5, len(y)) + np.random.randint(1, 1000, len(y)),
        "x2": y * np.random.uniform(0.5, 0.7, len(y)) + np.random.randint(1, 10000, len(y)),
    }
)
df["y"] = y
train, test = model_selection.train_test_split(df, train_size=150)

arima = pm.auto_arima(
    train["y"],
    train.drop(columns="y"),
    error_action="ignore",
    trace=True,
    suppress_warnings=True,
    maxiter=5,
    seasonal=True,
    m=12,
)

# preds1 takes the expected X args
preds1 = arima.predict_in_sample(X=train.drop(columns="y"))

# preds2 takes xargs with the correct dims, but different values from those used for preds1
preds2 = arima.predict_in_sample(X=train.drop(columns="y") + 1000)

# preds3 takes only x2, not x1, and x2 is subset to only 10 observations
preds3 = arima.predict_in_sample(X=train[:10].drop(columns=["y", "x1"]))

len(preds1)  # 150
len(preds2)  # 150
len(preds3)  # 150

all(preds1 == preds2)  # True
all(preds2 == preds3)  # True

arima.summary()  # To confirm that indeed x1 and x2 are in the model

Versions

System:
    python: 3.11.7 (main, Jan 24 2024, 16:45:17) [Clang 15.0.0 (clang-1500.1.0.2.5)]
executable: /Users/my-user/repos/forecasting/.venv/bin/python
   machine: macOS-14.2.1-arm64-arm-64bit

Python dependencies:
 setuptools: 69.0.3
        pip: 23.3.1
    sklearn: 1.4.0
statsmodels: 0.14.1
      numpy: 1.26.3
      scipy: 1.12.0
     Cython: 3.0.8
     pandas: 2.2.0
     joblib: 1.3.2
   pmdarima: 2.0.4

Expected Behavior

I expect the code to break if an X of incorrect dimensions is provided, and predictions to depend on the values of a correctly-dimensioned X.

Actual Behavior

The code does not break, and there is no difference in output.

Additional Context

No response