A jump-component portfolio earns a 0.61 Sharpe against 0.48 for the same machinery applied to raw returns, before costs. We could reproduce neither figure from the text. The settings needed to determine them remain unspecified.

Wei, Gao, Su and Zhang study five daily indices from 31 December 2001 to 25 October 2024: the S&P 500, the U.S. Dollar Index, the Bloomberg Commodity ex-Precious Metals Index, the GSCI Precious Metals Index and the S&P U.S. Treasury Futures Index. They pass each series through JOT (Jump, Oscillation, Trend), a variational filter from Huska et al. (2023) solved by ADMM. The filter fits a piecewise-constant part, a smooth part and a zero-mean high-oscillation part. In the authors' economic interpretation, trend reflects macro fundamentals, oscillation reflects sentiment and liquidity cycles, and jumps reflect crisis shocks. They argue that spillovers propagate differently through these channels, giving an allocator reason to separate them.

For each component, the authors estimate TVP-VAR connectedness using lag 4, selected by AIC, and a 100-step horizon. The total connectedness index (TCI) measures the average share of an asset's forecast-error variance attributable to the other four. The printed figures scale that average by 5/4 (see below). TCI reaches 37.12 for returns, 93.77 for trend, 38.78 for oscillation and 77.52 for jumps. Commodities lead the trend network with net +12.68; stocks alone transmit on net in oscillation (+5.58).

The portfolio Sharpe ratios are 0.61 for jump, 0.55 for trend, 0.48 for returns, 0.47 for oscillation and 0.25 for residual. The authors interpret the trend allocations through the Merrill Lynch Investment Clock: equities in 2003 to 2004, commodities in 2005 to 2007, bonds in late 2008.

Where the reconstruction stops

The paper consistently names four of the five instruments, including a futures index for the bond allocation. Precious metals changes identity: the data section names GSCI, while the robustness section names Bloomberg. The sample dates, lag and horizon are explicit. Our immediate obstacle is more specific. We hold no FX spot or futures prices, and the Dollar Index supplies one of only five network nodes.

In the returns network, the Dollar Index receives 33.51 of its variance from the other four. This is the largest entry in the returns panel's FROM column, which measures the share each asset receives from the others. Removing that node would change the network. We therefore did not attempt a four-asset version.

Which settings would an implementer use?

The filter is the first unresolved choice. The paper calls γ1, γ2 and γ3 "appropriately selected" without publishing their values. It also omits values for a, β and κ, giving only the condition β = τaγ1, τ > 1. These settings appear nowhere in the main text; an online supplement exists, but we have not seen it. The authors acknowledge that the ADMM design "does not necessarily guarantee global convergence of the overall non-convex problem." Different settings for a non-convex penalty can produce different decompositions.

The component labels create another problem. The paper labels v* as trend even though it is piecewise-constant, a step function of the sort associated with jumps. It labels n* as jump while describing it as zero-mean high oscillation that "can include additive white noise". The Huska et al. title puts Jump first, and the piecewise-constant term naturally carries jumps. Follow Huska et al.'s labelling convention and the resulting jump series becomes this paper's trend. A fourth residual series also appears in a three-component model. We suspect it is the fidelity misfit, f minus the three fitted parts, but cannot establish that from the text.

Even the input is ambiguous. The paper describes decomposing prices, then reports descriptive statistics for returns. We could not determine whether component levels or component changes enter the VAR.

Timing matters more. Their Eq. (2) optimises over a finite signal of length N, and we found no rolling or expanding re-estimation. Taken literally, prices after date t shape the component value at t. Weights derived from that value inherit the same dependence.

The paper then offers three possible portfolio constructions:

The weights table gives returns-panel means ranging from 0.1812 to 0.2243, summing to 1.0001. Means across five assets that sum to one are consistent with Eq. 44 or a five-asset optimiser. The pairwise construction in Eqs. 39 and 40 would not yield those means. Eq. (42) applies weights to "original returns," making the component portfolios tradable apart from the timing problem. Yet the cumulative-return figure uses component labels. The paper does not state how it annualises the Sharpe ratios.

Table 2 and the 5/4 adjustment

Eq. (29) defines TCI as the unadjusted average of FROM. The five returns entries in FROM average 29.70, whereas the table reports 37.12. Multiplying 29.70 by 5/4, the N/(N−1) correction, explains the difference. Trend follows the same calculation: its average of 75.01 times 1.25 yields 93.77. Once recognised, the adjustment is harmless, though the table still disagrees with the formula.

The residual discrepancy is harder to reconcile. The prose describes it as the lowest at 22.41%, but the table reports 75.45. Neither scaling connects those figures.

A 0.13 Sharpe gain, with weights ranging from 0 to 1

Jump improves on returns by 0.13; trend improves by 0.07. Oscillation loses 0.01. These comparisons come from five assets along one in-sample path. We found no test of the Sharpe differences.

The conclusion also claims an advantage in hedging effectiveness. Table 3 supports the opposite reading. For stocks, raw returns deliver 0.87, versus 0.72 for trend and 0.75 for jump. Bonds deliver 0.55 versus 0.07 and 0.15. Commodities and precious metals also decline, from 0.84 and 0.86 to 0.67 to 0.72. Currency falls from 0.17 to −0.72 and −0.57. Every p-value is printed as 0.00, including the negative ones.

Component weights vary much more than returns weights. Their standard deviations range from 0.0453 to 0.0636 for returns, 0.1886 to 0.2441 for trend and 0.1480 to 0.2051 for jump. Several trend and jump weights reach 1.0000. These figures describe dispersion in weight levels. A position held at 1 for a year before falling to 0 has a large standard deviation and two trades. Without reported turnover, the trading-cost drag remains unknown.

The authors explicitly call their results "benchmark evidence under a frictionless trading assumption," and the abstract limits the advantage to "in the pre-transaction-cost setting." This is the appropriate framing. The abstract also says "particularly in trend- or jump-dominated regimes," but we found no regime-conditional Sharpe figures supporting that claim.

The proxy exercise substitutes DJIA, 10-year T-note futures and COMEX gold. Total connectedness survives the substitution: trend reaches 95.22 versus the baseline 93.77, and jump reaches 84.96 versus 77.52. Network leadership changes. Commodity trend net moves from +12.68 to −2.33 despite the commodity index remaining unchanged. Precious-metals trend net shifts from −7.25 in the baseline table to +9.64 with the proxies.

The authors call the flows highly consistent. They also acknowledge that the proxies demonstrate "index-substitution resilience" and do not validate the framework in less liquid, emerging or non-U.S. markets. Even that narrower claim fails for the trend network.

We would take the 0.61 seriously if a trailing-window JOT with published parameters retained a gap over 0.48 after charging turnover.