Real-time analysis¶

Real-time forecasting uses the forecast-realtime package to refit a model at each historical information date. Each fit uses only observations available at that date, so the resulting forecast sequence avoids look-ahead bias.

This guide uses sampled data. It assigns a synthetic release date to each observation, then passes the resulting vintage table to forecast_realtime. Replace the sampled frames with a real-time database when applying the workflow to published data.

Install the integration packages before running the examples:

pip install -e ".[realtime]"

Generate sampled vintage data¶

sample_combo_data() returns a quarterly target and monthly and quarterly regressors. The code below adds a one-month publication lag to create a repeatable synthetic vintage schedule.

import pandas as pd

import forecast_evaluation as fe
import forecast_realtime as rt
import news_decomp as nd
from nowcast_midas.utils import sample_combo_data

target, regressors, _ = sample_combo_data(
    n_quarters=60,
    n_lags=6,
    seed=42,
)

target["vintage_date"] = target["date"] + pd.offsets.MonthEnd(1)
target["metric"] = "pop"
regressors["vintage_date"] = regressors["date"] + pd.offsets.MonthEnd(1)
regressors["metric"] = "pop"

outturns_data = pd.concat([target, regressors], ignore_index=True)
outturns_data["frequency"] = outturns_data["frequency"].replace({"QE": "Q", "ME": "M"})
data = fe.NowcastData(outturns_data=outturns_data)

The sampler uses QE and ME frequency codes. The real-time integration expects the equivalent Q and M codes, so the adapter translates them in the copied vintage table.

The returned frames use these columns:

Define the model and refit at each vintage¶

Define the same MIDAS specification for every vintage. The example forecasts the quarterly target from one monthly indicator.

ForecastMIDAS wraps MIDAS for the forecast_realtime.RealTimeModel runner. The runner filters the vintage table, refits the model, and stores each forecast under label.

model = rt.models.ForecastMIDAS(
    formula="quarterly_target ~ monthly_1",
    method="almon",
    n_lags=6,
    estimator="ols",
    horizons=[0, 1],
)
rt_model = rt.RealTimeModel(models=model, data=data)

rt_model.forecast(
    y_variables=["quarterly_target"],
    X_variables=["monthly_1"],
    data_transformation={
        "quarterly_target": "pop",
        "monthly_1": "pop",
    },
    label="sampled sc-midas",
    first_vintage=target["vintage_date"].iloc[24].strftime("%Y-%m-%d"),
    last_vintage=target["vintage_date"].iloc[-2].strftime("%Y-%m-%d"),
    reconstruct_levels=False,
    decomp=True,
)

The runner stores one forecast for the MIDAS model at each sampled information date. The date column identifies the forecast target, while vintage_date identifies the information date used for the fit.

Pass the resulting decomposition table to news_decomp for a news summary:

news = nd.NewsData(rt_model.decompositions)
news.summary()

Use the same sampled frames to compare model specifications, horizons, and combination methods without relying on external data or services.