Backtest Models¶

Backtesting runs a model across historical data vintages. At each vintage, RealTimeModel fits a fresh model, exposes only the information then available, and stores the forecasts in a shared ForecastData object.

1. Load data¶

Use ForecastData for one-frequency data or NowcastData when multiple vintages occur within a target period.

import forecast_evaluation as fe
import forecast_realtime as rt

data = fe.ForecastData(load_fer=True)

2. Configure a model¶

model = rt.models.ForecastOLS(
    label="ols",
    formula="cpisa ~ gdpkp + unemp",
)

The formula selects the target and regressors. Declare transformations in the vintage loop instead of applying them manually in the model.

3. Run the vintage loop¶

rt_model = rt.RealTimeModel(data=data, models=model)

rt_model.forecast(
    y_variables=["cpisa"],
    X_variables=["gdpkp", "unemp"],
    data_transformation={
        "cpisa": "pop",
        "gdpkp": "pop",
        "unemp": "levels",
    },
    step_frequency="Q",
    steps=8,
    first_vintage="2015-01-01",
    last_vintage="2020-12-31",
    label="ols",
)

Read forecasts from rt_model.data. When the model supports decomposition, forecast(..., decomp=True) also populates rt_model.decompositions.

4. Inspect the result¶

rt_model.data.summary()
rt_model.data.run_dashboard()

For autoregressive models, pass y_lags=4 or a model-specific lag argument. The real-time loop manages the expanding information set and recursive forecast horizons.