MIDAS Regression

Definition

MIDAS (Mixed Data Sampling) regression is a flexible distributed-lag model that relates a low-frequency dependent variable to a weighted sum of high-frequency predictors, with the lag weights parameterized by a small number of free parameters. The framework accommodates any mismatch between the frequency of the outcome and the frequency of the predictors without requiring aggregation of the predictors to the lower frequency first.

Key Ideas

Parsimonious weighting over many lags. The core problem in multi-frequency forecasting is that a lagged predictor at frequency $m$ requires $k^{\max}\cdot m$ coefficients. MIDAS constrains the lag polynomial to a parametric family (typically Beta or Almon), collapsing $k^{\max} m$ free parameters to 2–4.
Beta polynomial weights. The most common MIDAS weighting uses the Beta density evaluated on a grid $\{1/k^{\max}, \ldots, 1\}$ : $b_H(k;\theta_1,\theta_2) = \frac{f(k/k^{\max},\theta_1,\theta_2)}{\sum_{j=1}^{k^{\max}} f(j/k^{\max},\theta_1,\theta_2)}, \quad f(z,a,b) = \frac{z^{a-1}(1-z)^{b-1}}{B(a,b)}$ Only two shape parameters $[\theta_1,\theta_2]$ are estimated, yielding a normalised, non-negative weight profile that can be monotone-decaying ( $\theta_1=1$ , $\theta_2>1$ ), hump-shaped, or near-flat.
Mixed-frequency generality. MIDAS nests standard distributed-lag models when predictor and outcome are sampled at the same frequency. It extends naturally to intraday–daily, daily–monthly, and quarterly–annual data mismatches.
Nonlinear least squares estimation. Because weights are nonlinear in $\theta$ , estimation requires nonlinear least squares (NLS); the objective is the sum of squared forecast errors over the calibration window.

How It Works

MIDAS Specification for Volatility Forecasting (Ghysels-Santa-Clara-Valkanov [GSV] 2006)

The baseline MIDAS regression (eq. 1.1 in GSV 2006) is:

$V_{t+H,t}^{(Hm)} = \mu_H + \phi_H \sum_{k=0}^{k^{\max}} b_H(k;\theta)\,\tilde{X}_{t-k,t-k-1}^{(m)} + \varepsilon_{Ht}$

where $V_{t+H,t}^{(Hm)}$ is realized variance over the next $H$ days, $\tilde{X}^{(m)}$ is a daily volatility measure computed from $m$ -period intraday returns, and $b_H(k;\theta)$ are Beta polynomial weights with $\sum_k b_k = 1$ .

Five Predictor Specifications

GSV (2006) evaluate five choices of $\tilde{X}^{(m)}$ :

Predictor	Formula	Notes
Realized variance $\tilde{Q}^{(m)}_t$	$\sum_{j=1}^m r_{t-(j-1)/m}^2$	Converges to $IV_t$ (Integrated Variance) as $m\to\infty$ ; contaminated by jumps
Squared daily return $r_t^2$	$(p_t - p_{t-1})^2$	Noisiest proxy; daily frequency ( $m=1$ )
Absolute daily return $\|r_t\|$	$\|p_t - p_{t-1}\|$	Less noise than $r_t^2$
Daily range $[hi\text{-}lo]_t$	$(p_t^{hi} - p_t^{lo})^2/(4\ln 2)$	Garman-Klass estimator
Realized power $\tilde{P}^{(m)}_t$	$\sum_{j=1}^m \|r_{t-(j-1)/m}\|$	Barndorff-Nielsen–Shephard 2003/2004; excludes jump component

Andersen-Bollerslev-Diebold-Labys (ABDL) autoregressive fractionally integrated (ARFI)(5,d) Benchmark (eq. 1.10)

The long-memory benchmark against which MIDAS is evaluated is:

$(1-\textstyle\sum_{k=1}^5 b_A(k)L^k)(1-L)^d\,\log\tilde{Q}_{t+1,t}^{(m)} = \mu + \varepsilon_t, \quad d\approx 0.401$

where the $b_A(k)$ are unconstrained autoregressive (AR)(5) coefficients applied to fractionally-differenced log realized variance. This model captures long memory explicitly via the fractional difference operator.

Empirical Results (GSV 2006)

Data: Dow Jones index + 6 stocks (DIS, GE, JPM, XOM, MCD, HON); April 1993 – October 2003; 2,669 trading days; $m=80$ five-minute intervals per day.

Performance metric: mean squared error (MSE) ratio $MSE^{MIDAS}/MSE^{ABDL}$ — values below 1 mean MIDAS beats the long-memory benchmark.

Predictor	Horizon $H$	In-sample MSE/ABDL (level)	Out-of-sample
Realized power $\tilde{P}^{(m)}$	1–20 days	0.606–0.912	0.714–0.897
Daily range $[hi\text{-}lo]$	1–20 days	$\approx 0.7$ – $1.05$	$\approx 0.8$ – $1.0$
Realized variance $\tilde{Q}^{(m)}$	1–20 days	$\approx 0.85$ – $1.20$	$\approx 0.9$ – $1.1$

Predictor ranking (consistent across assets and horizons): $\tilde{P}^{(m)} \succ [hi\text{-}lo] \succ \tilde{Q}^{(m)} \succ |r| \succ r^2$

Why realized power wins: The Beta MIDAS with realized power places weight $b(1) = 38.3\%$ on yesterday's power at $H=1$ , declining to $8.6\%$ at $H=20$ days (Table 2). The remaining weight spreads over ~30 lags. The ABDL ARFI model effectively requires $>50$ lags to represent its long-memory tail, making it parameter-heavier despite the long-memory shortcut.

No gain from direct intraday use: Replacing daily $\tilde{P}^{(m)}$ with raw 5-minute absolute returns as MIDAS regressors (eqs. 3.1–3.2) yields no improvement in MSE. The information in 5-min data is fully captured by daily realized power.

Why Realized Power Has Higher Persistence and Forecasting Power

Under the Barndorff-Nielsen–Shephard (2004) framework, realized power converges to the continuous component of quadratic variation (integrated variance), while realized variance $\tilde{Q}^{(m)}$ includes jumps. Since jumps are transient, their inclusion in the predictor reduces persistence and degrades multi-step forecast accuracy. Empirically, $\log\tilde{P}^{(m)}$ has a higher AR(1) coefficient and smoother autocorrelation function (ACF) decay than $\log\tilde{Q}^{(m)}$ .

Why It Matters

Mixed-frequency data fusion. MIDAS provides a principled framework for combining series sampled at different rates without the information loss of pre-aggregation or the parameter explosion of unconstrained distributed lags.
Jump exclusion as a design principle. The realized power result suggests that for multi-step volatility forecasting, removing the jump component of quadratic variation is more valuable than maximizing instantaneous precision. This guides predictor choice in applied forecasting.
Alternative to long-memory models. MIDAS with Beta weights and ~50 lags replicates or beats autoregressive fractionally integrated moving average (ARFIMA)( $d\approx0.4$ ) performance without requiring fractional integration theory, simplifying inference and extending naturally to multivariate settings.

Open Questions

Optimal $k^{\max}$ selection. The choice of maximum lag length is informal; a formal cross-validation or information criterion for $k^{\max}$ in MIDAS models is not established.
Multivariate MIDAS. The univariate framework does not address covariance forecasting for portfolios; extending to realized covariance matrices raises dimensionality challenges.
Microstructure noise in direct intraday MIDAS. The null finding on direct 5-min regressors may reflect bid-ask noise contaminating raw returns; noise-robust realized measures (Zhang-Mykland-Aït-Sahalia 2005) might restore a high-frequency advantage.