Fix Statsmodels future forecast horizon

[Li Jiang (thinkall)]

Why are these changes needed?

Statsmodels 0.15 can interpret future datetime boundaries as positional indices when an ARIMA/SARIMAX model does not retain a supported date index. In test_numpy_large, that turned a 10-row validation horizon into 59,990 out-of-sample steps and rejected the 10-row exogenous feature matrix.

Use step-based forecast() for prediction frames that begin after the estimator's training end date. Date-based predict() remains unchanged for in-sample predictions, and the shared forecast() API preserves compatibility with ARIMA, SARIMAX, Holt-Winters, and simple forecasters.

Failed job: https://github.com/microsoft/FLAML/actions/runs/35058894402/job/104684468013

Related issue number

N/A

Checks

• I've used pre-commit to lint the changes in this PR (note the same in integrated in our CI checks).
• I've included any doc changes needed for https://microsoft.github.io/FLAML/. See https://microsoft.github.io/FLAML/docs/Contribute#documentation to build and test documentation locally. (No documentation changes needed.)
• I've added tests (if relevant) corresponding to the changes introduced in this PR. (test_numpy_large is the existing regression coverage.)
• I've made sure all auto checks have passed.

Targeted validation:

python -m pytest test\automl\test_forecast.py::test_numpy_large test\automl\test_forecast.py::test_models test\automl\test_forecast.py::test_numpy test\automl\test_forecast.py::test_log_training_metric_ts_models -q --reruns 2 --reruns-delay 1

Result: 3 passed, 1 skipped.

[Li Jiang (thinkall)]

Self-authored PR, so GitHub does not permit me to approve or request changes. One blocking issue found:

• flaml/automl/time_series/ts_model.py:358-362: the new path ignores the requested timestamps and always returns the first X.shape[0] periods immediately after training. If the requested future frame begins after a gap, it silently returns forecasts for the wrong dates. Preserve date-based predict() for delayed starts, or forecast through the skipped offset and slice; exogenous models must also account for every intervening step. Add a regression test for a noncontiguous/delayed future start across supported index types.

[Copilot]

🟡 Changes recommended

The future-horizon boundary check must use the model’s training end date.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

This PR updates Statsmodels forecasting to avoid incorrect future horizons in Statsmodels 0.15.

Changes:

• Uses step-based forecast() for future prediction frames.
• Preserves date-based predict() for other predictions.
• Reuses preprocessed exogenous features.

File summaries

File	Summary
flaml/automl/time_series/ts_model.py	Adds future-horizon forecasting logic. Moderate issue (3 votes): the boundary check uses the dataset end date instead of the model training end date, so the fix may be bypassed for holdout/CV data.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.