MIDAS and Mixed-Frequency Regression

You want to forecast next quarter's GDP growth using daily credit-spread data. GDP is measured once a quarter — roughly 65 trading days per quarter's worth of daily spread observations sit between each GDP release. The naive fix is to average the daily spreads down to a single quarterly number, throwing away 64 out of every 65 daily observations' worth of timing information. The other naive fix is to include all 65 daily lags as separate regressors, which would need more parameters than you have quarters of history to estimate them with. MIDAS (Mixed Data Sampling) regression is the middle path: it keeps every daily lag in the model, but forces their coefficients to follow a smooth, low-dimensional weighting pattern instead of estimating each one freely.

An analogy: how much should yesterday matter versus a month ago

A weather forecaster predicting this week's average temperature naturally weights yesterday's reading more than a reading from three weeks ago — not by discarding the older data, and not by treating every day as equally informative either, but by applying a smoothly decaying weight that gives recent days more influence. MIDAS applies exactly this logic to economic forecasting: instead of estimating 65 separate, noisy coefficients (one per daily lag, an impossible task with limited quarterly history), it assumes the coefficients follow a smooth decay curve controlled by just two or three parameters, letting recent observations count more than older ones without discarding any of them.

The idea, one symbol at a time

A MIDAS regression forecasting a low-frequency target $y_t$ using a high-frequency predictor $x_{t,k}$ (the $k$-th high-frequency observation within low-frequency period $t$) takes the form

$$y_t = \beta_0 + \beta_1 \sum_{k=1}^{K} w(k; \theta) \, x_{t,k} + \varepsilon_t,$$

where $w(k; \theta)$ is a weighting function — commonly an exponential Almon lag polynomial — controlled by just a small number of parameters $\theta$, and normalized so the weights sum to 1. In plain English: instead of a separate coefficient on each of the $K$ high-frequency lags, all $K$ lags share a single overall coefficient $\beta_1$, and the relative importance of lag $k$ versus lag $k'$ is set by the shape of $w(\cdot;\theta)$, which typically declines smoothly as $k$ moves further back in time. This collapses what would be $K$ free parameters down to just $\theta$ (often only 2 numbers) plus $\beta_1$, making estimation feasible even with a modest number of low-frequency observations.

Worked example 1: exponential weights by hand

Suppose a simplified MIDAS setup uses only the last 3 daily lags within the quarter (illustrative — a real MIDAS model typically uses far more) with an exponential weighting $w(k) \propto e^{-\theta k}$, $\theta = 0.5$, for $k=1,2,3$ (1 = most recent). Unnormalized weights: $e^{-0.5(1)}=0.607$, $e^{-0.5(2)}=0.368$, $e^{-0.5(3)}=0.223$; summing to $1.198$. Normalized: $0.507, 0.307, 0.186$. If the last 3 daily credit spread readings were $2.1, 1.9, 1.7$ (bps, most recent first), the MIDAS-weighted predictor value is

$$0.507(2.1) + 0.307(1.9) + 0.186(1.7) \approx 1.065 + 0.583 + 0.316 = 1.964.$$

This single number, $1.964$, then enters the regression as $\sum_k w(k)x_{t,k}$, with the most recent reading (2.1) contributing more than half the total weight despite being just one of three lags.

Worked example 2: what a flatter theta changes

Now suppose cross-validation instead selects a much smaller $\theta = 0.1$ (slower decay, near-equal weighting). Unnormalized weights: $e^{-0.1}=0.905$, $e^{-0.2}=0.819$, $e^{-0.3}=0.741$, summing to $2.465$; normalized: $0.367, 0.332, 0.301$ — close to an equal split. The weighted predictor becomes $0.367(2.1)+0.332(1.9)+0.301(1.7)\approx1.914$, barely different from an unweighted average since the weights are nearly flat. $\theta$ matters far more when recent and older lags diverge sharply — a crisis quarter with a sudden spread spike is exactly where a flat-weighted average would badly underweight the most relevant, most recent signal.

Drag the decay parameter in the explorer above — it traces the same shape MIDAS uses to weight high-frequency lags, faster decay concentrating weight on the most recent observations, slower decay spreading weight more evenly across the whole window.

What this means in practice

MIDAS regression is used for nowcasting — updating a GDP, inflation, or earnings forecast in real time as new daily or weekly data arrives, well before the official quarterly figure is released — and for volatility forecasting, where realized daily volatility is used to predict a lower-frequency (say, monthly) risk measure. It sits between the two lazy extremes of "aggregate everything to the lowest frequency" and "estimate a hopeless number of separate lag coefficients," and is the standard tool whenever your predictor and target genuinely live at different sampling frequencies.