A risk desk should care that Tensor-BEKK ranks fourth of six model-based estimators on realized volatility. Its 45-cell covariance forecast is inexpensive to run and straightforward to code, assembled from 2x2 mode factors for the intercept, the ARCH and the GARCH matrices. The open question is what risk forecasting gives up when covariance dynamics are forced into a separable form. Gong and Jiang's own out-of-sample table suggests the cost is small. Any benefit is equally hard to see.
The portfolio grid enters the covariance model
Arrange returns as an order-m array, vectorize them to x_t with d = prod d_k, then use the classical BEKK(1,1) recursion H_t = Ω + A x_{t-1} x_{t-1}' A' + B H_{t-1} B'. The authors constrain Ω, A and B to a single Kronecker product across modes. Thus A = A_m ⊗... ⊗ A_1, with each A_k of size d_k by d_k. The underlying recursion stays intact, preserving positive definiteness, while the parameter count drops from order d^2 to order sum_k d_k^2. Their 2x2 simulation design has 19 free parameters against 42 for unrestricted BEKK(1,1).
One identity carries the economic interpretation. The (i_1..i_m),(j_1..j_m) entry of A is the product over modes of (A_k)_{i_k j_k}. A shock from the small-cap resources cell can reach the large-cap consumer cell through one factor from each mode: a sector term times a size term times a growth term. This construction makes every cross-mode interaction separable. A_k governs shock propagation along mode k, while B_k governs the propagation of past conditional covariance.
Strict stationarity and ergodicity have a sufficient condition: rho(A⊗A + B⊗B) < 1, provided the innovations have a density that is positive near zero. Identification relies on Cholesky and (1,1)-entry normalizations, together with an injectivity condition on the global BEKK map. Gaussian QMLE is consistent and asymptotically normal with a sandwich covariance.
The restricted score tests cover off(A_k) = 0 and off(B_k) = 0. Each has d_k(d_k - 1) degrees of freedom, rising to 2d_k(d_k - 1) jointly. A portmanteau test uses the quadratic-form residual u_t = x_t' H_t^{-1} x_t - d.
For larger panels, TF-BEKK applies a T-BEKK recursion to a low-dimensional Tucker factor tensor. Step one uses contemporaneous modewise PCA. The authors choose it because the factor process is a martingale difference sequence whose nonzero-lag autocovariances vanish, making the usual autocovariance-based tensor factor estimators inapplicable here. Step two runs T-BEKK QMLE on the recovered factors.
The paper has two applications. The first is a 2x2 currency futures matrix built from 1,335 weekly VAR(5) residuals spanning 2001-02-09 to 2026-09-04. The second contains 45 Chinese A-share cell portfolios: five aggregated sectors crossed with three size groups and three groups formed on growth in total assets. These cells come from a fixed 175-stock universe and hold three or four stocks apiece.
Estimation runs through an 800-day rolling window. The out-of-sample period contains 250 trading days from 2024-04-09, with re-estimation every 50 days and daily covariance forecast updates. Neither universe is available to us as studied. The currency and A-share figures therefore do not transfer to a US implementation; we are running the mechanism on US sector-by-market-cap cells instead.
What the 250-day results actually show
TF-BEKK records a 22.21% annualized return, 18.70% annualized volatility and an information ratio of 1.19. Those are the best AV and IR among the seven methods. A long-only global minimum-variance objective should be judged by the volatility column, where TF-BEKK comes fourth among the six model-based estimators: MF-GARCH 18.05, scalar BEKK 18.31, GO-GARCH 18.50, TF-BEKK 18.70, RiskMetrics 18.87, TDCC 19.61. Equal weighting comes in at 22.86.
Every covariance model reduces variance, with only 1.56 percentage points separating them.
The authors acknowledge that MF-GARCH produces the lowest SD, at 18.05%, against 18.70% for TF-BEKK. They answer with TF-BEKK's highest cumulative return at the end of the evaluation period, alongside its leading AV and IR. Its IR advantage over MF-GARCH, 1.19 against 0.90, comes from a one-year return gap of 22.21% against 16.23% under no-short and full-investment constraints. Returns create the gap. TF-BEKK's realized volatility remains in the middle of the pack.
The simulations and two applications, the authors write, "illustrate the finite-sample behavior and empirical usefulness" of the methods. They are equally direct about how the comparison treats the cell portfolios: frictionlessly rebalanced return indices. Performance is "measured before transaction costs and with unrestricted rebalancing during stock suspensions."
Table 5 gives AV, SD and IR. We did not find turnover, drawdown, or a test of the IR differences. The universe is selected using average log market cap through 2024-04-08 and then held fixed, which keeps the selection from reaching into the evaluation period. The training portfolios, however, use names known to have survived through the end of the training window.
Choices left to the implementer
Rank comes first. Every estimation window fixes the factor rank at (2,2,2). The theory assumes the ranks are given, and the paper supplies no selection rule. With r = 8, the second-step likelihood inverts 8x8 matrices instead of 45x45, which provides the extension's computational gain. Choosing rank separately in each window creates a different model. The conclusion leaves time-varying factor structures, including evolving loading spaces and factor ranks, as open work.
Orientation creates another decision. The reference matrices J_k, which identify the latent basis, are prespecified under a condition in the supplement. Orientation cancels when the sole output is the reconstructed covariance forecast. It matters for mode-specific spillover tests on the latent factors, leaving the selection of J_k to the implementer.
Sigma_e is time-invariant and estimated through nonlinear shrinkage, with further details placed in the supplement. We therefore used rolling-window residual shrinkage within each 800-day window. Initialization is fully specified as x_0 = 0, H_0 = I_d.
Mode selection receives no rule. In the application, the authors fix sector, size and asset-growth buckets, leaving the implementer to decide what should become a mode. The chosen characteristics and their bucket widths determine which covariance patterns the model can represent. We encountered the same kind of ambiguity when reproducing Adaptive LASSO-MGARCH, where reproducibility depended on one estimation step (note).
Inference carries less weight at these dimensions
The asymptotic treatment holds mode dimensions fixed while T grows. TF-BEKK oracle equivalence requires sqrt(T)(delta_T + d^{-1/2}) to vanish, where delta_T = max_k {(T d_{-k})^{-1/2} + d_k^{-1}}. Substitute 5x3x3 and T = 800. The d_k^{-1} term by itself is 1/3, d^{-1/2} is 0.149, and the resulting product is roughly 14. Covariance forecasting does not depend on that theorem. The second-step standard errors do.
The currency application says more than the allocation exercise, largely through a negative result. Estimated persistence reaches 0.99556 near the stationarity boundary. The GARCH diagonals are 0.947 (SE 0.012), 0.973 (0.008) and 1.002 (0.010).
All three mode-1 spillover tests come up empty: 0.043 (p = 0.979), 0.200 (p = 0.905) and 0.859 (p = 0.930). Off-diagonals contribute 2.93% of the squared ARCH norm and 0.012% of the GARCH norm. At 5%, the portmanteau statistic does not reject. Q(6) = 11.819 (p = 0.066) is the nearest call. The machinery behaves, yet there is no cross-currency-group transmission to detect.
Kronecker separability is imposed without a corresponding test. The portmanteau diagnostic looks for serial dependence in quadratic-form residuals, while its simulations address omitted lag-2 ARCH and GARCH terms only. We did not find a separability test against either unrestricted BEKK or Kronecker-sum BEKK.
Evidence that would change my view is straightforward: TF-BEKK delivering the lowest realized volatility, net of turnover, on US sector-by-size cells across several multi-year windows using identical cell construction. Realized volatility is the column this estimator is supposed to win.