Combine Forecasts¶

forecast_combo combines forecasts stored in a ForecastData object. Add the individual model forecasts before fitting a combination.

1. Produce individual forecasts¶

Any compatible model workflow can supply the individual sources. For example, an OLS and MIDAS model can share a NowcastData object:

import forecast_combo as fc

forecast_data = rt_model.data
combo = fc.ForecastCombo(forecast_data=forecast_data)

2. Fit a combination¶

combo.fit(
    sources=["ols", "midas"],
    variables=["quarterly_a"],
    method="rmse",
    metric="levels",
    label="rmse combo",
)

Available methods include average, rmse, mse, mae, huber, least_squares, and constrained_least_squares. average assigns equal weights; error-based methods estimate weights from historical forecast errors.

Use training_start, training_end, window_size, discount_param, period_filter, and k to control the fitting period and outturn maturity.

3. Inspect weights and forecasts¶

weights = combo.weights
combined_data = combo.forecast_data

fc.heatmap_by_vintage(weights, method="rmse", variable="quarterly_a")
combo.run_forecast_dashboard()

combo.forecast_data includes the combined forecast. Evaluate it alongside the component forecasts without rebuilding the data object.

Hierarchical combinations¶

Use ComboSpec when one combination supplies another combination's source:

from forecast_combo import ComboSpec

first_stage = ComboSpec(
    name="model_average",
    sources=["ols", "midas"],
    method="average",
)

combo.fit(
    sources=ComboSpec(
        name="top_level",
        sources=[first_stage, "benchmark"],
        method="rmse",
    ),
    variables=["quarterly_a"],
)