Mean Square Errors of factors extracted using principal components, linear projections, and Kalman filter

Factor extraction from systems of variables with a large cross-sectional dimension, $N$, is often based on either Principal Components (PC)-based procedures, or Kalman filter (KF)-based procedures. Measuring the uncertainty of the extracted factors is important when, for example, they have a direct interpretation and/or they are used to summarized the information in a large number of potential predictors. In this paper, we compare the finite $N$ mean square errors (MSEs) of PC and KF factors extracted under different structures of the idiosyncratic cross-correlations. We show that the MSEs of PC-based factors, implicitly based on treating the true underlying factors as deterministic, are larger than the corresponding MSEs of KF factors, obtained by treating the true factors as either serially independent or autocorrelated random variables. We also study and compare the MSEs of PC and KF factors estimated when the idiosyncratic components are wrongly considered as if they were cross-sectionally homoscedastic and/or uncorrelated. The relevance of the results for the construction of confidence intervals for the factors are illustrated with simulated data.

💡 Research Summary

This paper investigates the finite‑sample mean‑square error (MSE) properties of three popular factor‑extraction techniques—Principal Components (PC), Linear Projection (LP), and Kalman Filter (KF)—in large‑dimensional dynamic factor models (DFMs). The authors consider a standard DFM where the N‑dimensional observation vector Yₜ is decomposed as Yₜ = ΛFₜ + εₜ, with Λ the loading matrix, Fₜ the r‑dimensional common factor (r ≪ N), and εₜ the idiosyncratic component. The loading matrix and the number of factors are assumed known to isolate the effect of the extraction method on factor uncertainty.

Two alternative conceptualizations of the factors are examined. In the “deterministic” case the factors are treated as fixed unknown constants; this is the implicit assumption underlying PC‑based estimators, which can be viewed as least‑squares (LS) estimators. In the “stochastic” case the factors are random variables. If they are i.i.d. the LP estimator coincides with a static linear projection of Yₜ onto the space spanned by Λ. If the factors follow a state‑space dynamics (e.g., AR(1)), the KF estimator provides the optimal conditional mean given past observations.

A central contribution is the systematic comparison of MSE matrices under three specifications of the idiosyncratic covariance Σε: (i) the full (potentially heteroscedastic and cross‑sectionally correlated) matrix, (ii) a diagonal matrix (cross‑sectional independence), and (iii) a spherical matrix (homoscedastic and independent). The authors derive closed‑form expressions for the MSE of each estimator under each Σε specification, assuming √N‑consistency (i.e., √N/T → 0 as N,T → ∞).

Key theoretical findings are:

1. When the true Σε is used, the MSE of all three estimators is diagonal, implying no spurious correlation among estimated factors.
2. Mis‑specifying Σε as diagonal or spherical introduces non‑zero off‑diagonal elements, creating artificial factor correlations.
3. For any given Σε, the LP estimator dominates the LS (PC) estimator: the difference MSE(LP) – MSE(LS) is negative semidefinite. This reflects the gain from treating factors as random and estimating their conditional means rather than a single realization.
4. The advantage of KF over LP is modest unless the factor persistence (the AR coefficient) is large. When persistence is high, KF’s use of the full state‑space structure yields a noticeably smaller MSE.
5. All nine combinations (3 Σε structures × 3 estimators) exhibit MSE that shrinks at rate O(N⁻¹); consequently the estimators are √N‑consistent and asymptotically equivalent after scaling by N.

Monte‑Carlo simulations corroborate the analytical results. The authors generate data with varying N, T, factor persistence, and idiosyncratic structures. The simulations show that (i) PC’s MSE inflates dramatically when Σε is heteroscedastic and cross‑correlated, because PC cannot correctly weight the observations without the full Σε; (ii) LP improves upon PC even when Σε is only diagonal; (iii) KF provides the lowest MSE when the factor follows a persistent AR process, especially for moderate N where the asymptotic equivalence has not yet set in.

From a practical standpoint, the paper argues that when factors are used directly as economic or financial indices—e.g., business‑cycle indicators, systemic‑risk measures, or scenario‑generation tools—accurate quantification of factor uncertainty is essential. In such contexts, the KF‑based extraction is recommended, particularly when (a) the cross‑sectional dimension N is small or moderate, (b) idiosyncratic heteroscedasticity is present, and (c) the underlying factors exhibit strong persistence. While estimating the full Σε is computationally demanding, the results suggest that even a diagonal approximation combined with KF yields superior performance to PC.

In summary, the paper provides the first comprehensive finite‑sample MSE analysis of PC, LP, and KF factor estimators, highlighting the trade‑offs between deterministic versus stochastic factor treatments and the importance of correctly modeling idiosyncratic covariance. The authors’ recommendations—use KF with heteroscedastic idiosyncratic modeling for small‑to‑moderate N and highly persistent factors—offer clear guidance for empirical researchers working with high‑dimensional macro‑financial datasets.