A trader holding minimum-variance stocks gets no reason from this paper's walk-forward test to switch once costs reach 5 bps. The stronger finding concerns how to combine portfolios. Spelta, Pecora and Maggi find that an optimal-transport barycenter beats a linear blend of the same regime portfolios. Its appeal depends on what you would otherwise hold.

How the barycenter allocates

The proposed allocation moves between bull, bear and neutral portfolios as sentiment changes. Its intended source of profit is a tilt toward stocks during greed and bonds during fear.

Each portfolio becomes a metric-measure space: portfolio weights supply the measure, while a similarity matrix derived from return correlations supplies the geometry, S_ij = 2 − √(2(1 − C_ij)). The authors begin with three Markowitz portfolios. Their universe contains 27 individual stocks, 11 bond ETFs and 5 inverse ETFs, including SDS and SH.

CNN Fear & Greed Index supplies the regime signal. Its seven market indicators include safe-haven demand, volatility, junk bond demand, price strength, breadth and put/call options. The resulting score runs from 0-100, with 0 denoting extreme fear and 100 extreme greed. The authors compare the index with a short and a long Kaufman adaptive moving average. A reading above both means bull; below both means bear. Everything else is neutral. The mixing vectors follow the regime:

Applying that mix directly to the input weight vectors produces the FG-portfolio, the paper's linear sentiment blend.

The GW-portfolio instead uses the mix as barycentric weights, which determine how much each input counts in the optimization. It chooses the allocation that minimizes weighted entropic Gromov-Wasserstein distortion from the three inputs. Block coordinate descent with Sinkhorn steps solves the problem, beginning at the FG weights. The entropy parameter epsilon controls concentration. In the authors' 10-asset simulation, epsilon 0.02 concentrates everything in one asset; epsilon 2 produces equal weights.

There is no return forecast. Markowitz portfolios fitted on trailing windows provide the inputs. Any edge must come from shifting toward equities in greed and duration in fear, while the barycenter smooths those shifts. Daily Yahoo prices cover January 2011 to March 2024, and the walk-forward test runs from 2012 to 2024. Each step selects epsilon from {0.02, 0.04, 0.1} and chooses among 27 mixing vectors using in-sample P&L. Including window lengths brings the full grid to 1296 paths.

Which benchmark would you hold?

GW versus FG is the clean comparison. Both use identical signals, inputs and mixing vectors; the aggregation method alone changes. The barycenter wins clearly. Full-sample Sharpe at 1 bp is 0.373 for GW versus 0.069 for FG. With costs at 5 bps, GW returns −0.19% and FG returns −26.80%. During 2022 at 1 bp, the losses were 20.03% for GW and 42.36% for FG.

That establishes a real improvement over linear regime blending.

For a trader, the stock minimum-variance portfolio is the more relevant alternative. It returned 35.95% with Sharpe 0.268 at 1 bp; GW returned 32.23% with Sharpe 0.373. GW wins only on Sharpe. Raise costs to 5 bps and stocks retain 28.48% with Sharpe 0.212, while GW drops to −0.19% with Sharpe −0.002.

The abstract claims superior risk-adjusted performance across the full evaluation window without specifying a cost level. It also describes year-by-year variation as "as expected from a dynamic strategy". In the body, superiority is restricted to 1 bp, and the authors acknowledge that 5 bps frictions hurt every active strategy. But −0.19% against the stock book's 28.48% is a full-window failure at 5 bps. Variation between years cannot explain it. Why leave the cost qualification out of the abstract? For the 5 bps run, the authors emphasize "turnover mitigation" against FG. That comparison holds, though it uses the weaker benchmark.

The cost convention matters. Net return equals gross return less c times a trade indicator: every rebalance incurs the same c, regardless of how much weight changes. Appendix F gives 69.3 rebalancing events per year, occurring on 27.5% of days, with a median interval of 2 days, for "the strategy". Yet the table's caption identifies the FG strategy. Whose turnover those figures describe is therefore harder to establish.

Adding 4 bps per event removes about 32 points from GW's cumulative return. Stocks lose 7.47 points under the same change. The paper does not let us determine whether charging in proportion to traded notional would cost more or less. We would want that cost run before relying on either figure.

Appendix F also mentions an entropy-based sentiment trigger whose definition we did not find in the methodology.

Protection in 2022, a larger drawdown overall

The barycenter does provide protection in 2022. At 1 bp, its drawdown was 22.21, compared with 29.67 for stocks and 43.09 for FG.

Across the full path, however, GW suffers a larger maximum drawdown than stocks. It remains below FG's 54.64 at 1 bp and 71.25 at 5 bps. GW's figures are 36.40 versus stocks' 33.87 at 1 bp, then 48.35 versus 35.08 at 5 bps. These are arithmetic measures. The paper adds daily returns rather than compounding them, then calculates drawdown on the resulting path. Interpret the figures as point losses on a running sum of daily returns.

The weak years are scattered through the sample. At 1 bp, losses were 5.66% in 2015 and 3.22% in 2018. The stronger years at 1 bp were 2013, 2014, 2019 and 2024, all with Sharpe above 2.1.

Before putting money behind GW

The headline P&L chart's red line comes from ex-post optimization. The authors acknowledge that this can leave results affected by overfitting. They defend the exercise: "However, this step is instrumental in revealing how the GW-portfolio behaves while varying parameter configurations." This review uses the walk-forward figures, which were not selected ex-post.

Appendix D also reports in-sample results, without an equivalent warning from the authors. GW-ExPost returns 74.10% with Sharpe 0.675 there. Replacing Fear & Greed with VIX raises the return to 137.03% and Sharpe to 0.785. The authors dismiss the VIX rule because of its volatility, 14.20% versus 8.93%, and greater tail risk. Sharpe already penalizes that volatility.

Hansen's SPA and White's Reality Check, tests for data snooping, then take GW-ExPost as their reference strategy. According to the paper, the global null hypothesis that no strategy outperforms GW Ex post cannot be rejected. The authors describe this as "statistical and practical dominance" of "the proposed barycentric allocation framework". Non-rejection cannot support that claim. The credit also goes to the entire framework, although the tested strategy is the tuned GW-ExPost.

The authors' DCC-GARCH simulation complicates the interpretability argument too. At epsilon 0.2, the barycenter loads assets 3 and 5 under all three mixing vectors tested. Asset 3 has no weight in either input. Geometry can therefore override the regime weights, making it harder to explain an allocation through its regime.

We could not run this ourselves. We lack the historical CNN Fear & Greed series, which the paper requires as its regime input. A VIX substitution would test a different signal. The paper's VIX variant returned 137.03% cumulatively with Sharpe 0.785 in-sample, compared with GW's 40.02% and 0.483 at 1 bp.

A walk-forward test charging costs against traded notional would change our view if GW kept its Sharpe above the stock book's 0.212 at 5 bps. Until then, the barycenter improves on averaging regime portfolios. At 5 bps per rebalance, it offers no improvement over holding minimum-variance stocks, with Sharpe −0.002 against 0.212.