Staggered-Combination MIDAS

Mixed-frequency nowcasting and short-horizon forecasting of quarterly targets (e.g. GDP) from monthly indicators, with forecast combination.

Experimental Python implementation of Moreira (2025)

Read the user manual for the full guide.

Features

• MIDAS regression with Almon, exponential-Almon, Beta and unrestricted lag polynomials.
• Quarterly OLS counterpart (OLS / OLSSpec) for hard quarterly indicators, sharing the same forecasting and dummy interface.
• Direct multi-horizon forecasting: one model per horizon h, fit and stored together.
• Outlier dummies on the target (dummy_periods) absorbed at the target frequency.
• End-to-end combination pipeline MidasCombo:
  • Average, inverse-error (mae, mse, rmse) and regression weights across indicators.
  • Rolling-window and discounted-error variants.
  • Combinations of combinations (a ComboSpec can reference other ComboSpecs as sources).
• NLS estimation of non-linear weight schemes (Almon-exp, Beta).

Project Structure

├── src/nowcast_midas/    # Source code
├── docs/            # Zensical documentation site
├── examples/        # Example scripts
├── tests/midas/     # Unit tests
└── ...

Installation

From PyPi:

pip install nowcast-midas

Dev version:

git clone https://github.com/bank-of-england/nowcast-midas.git
cd nowcast-midas
pip install -e .                      # runtime
pip install -e ".[dev]"               # + test / lint / docs tooling
pip install -e ".[realtime]"          # + real-time (vintage) analysis stack

Python ≥ 3.10.

Quick start

from nowcast_midas import ComboSpec, MidasCombo, MidasSpec, OLSSpec
from nowcast_midas.utils import sample_combo_data

# Simulated mixed-frequency data: three monthly series, one quarterly
# regressor, one quarterly target, plus one injected outlier.
target_df, regressors_df, info = sample_combo_data(n_quarters=60, seed=42)
outlier = info["outlier_date"]

midas_monthly_1 = MidasSpec(
    "monthly_1", method="almon", n_lags=6, dummy_periods=[outlier]
)
midas_monthly_2 = MidasSpec(
    "monthly_2", method="almon", n_lags=6, dummy_periods=[outlier]
)
midas_monthly_3 = MidasSpec(
    "monthly_3", method="unrestricted", n_lags=3, dummy_periods=[outlier]
)
ols_quarterly_1 = OLSSpec("quarterly_1", n_lags=1, dummy_periods=[outlier])

soft = ComboSpec(
    "soft",
    sources=[midas_monthly_1, midas_monthly_2, midas_monthly_3],
    method="mse",  # inverse mean-squared-error weights — see docs/methods/combo.md
    window=8,
    discount_rate=0.95,
)
final = ComboSpec("final", sources=[soft, ols_quarterly_1], method="regression")

model = MidasCombo(combo_specs=final, horizons=3)  # horizons is a COUNT, not an index
model.fit(target=target_df, regressors=regressors_df)

oos = model.forecast()  # long format: one row per (spec, horizon step);
# columns horizon, spec, value, date
print(oos.head())
model.summary(horizon=0)  # prints and returns the text

This is the same example as docs/index.md; runnable end-to-end scripts are in examples/ and rendered in Worked examples.

Selected documentation

• SC-MIDAS framework
• Weighting schemes
• MIDAS model
• SC-MIDAS pipeline
• Real-time analysis

Selected examples

• MIDAS, MultiMIDAS, and SC-MIDAS
• SC-MIDAS with forecast-realtime

Contributing

• See CONTRIBUTING.md for details on how to contribute to the code.
• Open an issue with questions or ideas.

Main references

• Moreira (2025)
• Ghysels, Santa-Clara and Valkanov (2004)
• Ghysels, Sinko and Valkanov (2007)

Data Classification

Bank of England Data Classification: OFFICIAL BLUE