Supriadi's sector rotation earns 276 bps a year over Plain NCO, yet leaves 295 bps on the table against SPY. Over August 2020 to December 2025, the model compounded at 13.41% gross. Plain NCO returned 10.65% and SPY 16.36%. Its total return reached 97.74%, versus 127.26% for SPY. Monthly rebalancing of an equal-weight portfolio of the 11 sectors earned 14.28%, with a Sharpe of 0.78 against the model's 0.63. At 17.71%, the model also had the highest annualized volatility among the strategies tested.
Supriadi acknowledges the comparison with SPY. Against the passive S&P 500, the model's Sharpe is 0.63 vs 0.86 and its Sortino is 1.07 vs 1.47. He writes: "Its edge is therefore not volatility-scaled." His case rests instead on maximum drawdown, Calmar ratio and turnover. The model's max drawdown was -14.22%, compared with -23.93% for SPY and -18.32% for 1/N. Its Calmar of 0.94 leads the table, although plain minimum variance suffered a smaller drawdown at -13.52%.
Inside Asym-Sticky-NCO
Nested Clustered Optimization (NCO) is López de Prado's method for avoiding a badly conditioned covariance inverse. It groups assets using a correlation distance, optimizes within each cluster, then allocates among the clusters as though each were an asset. The paper's Plain NCO applies this procedure to ordinary symmetric covariance.
Supriadi objects that symmetric variance treats a sector rally as risk. During a bull market, variance-minimizing NCO consequently shies away from the technology and consumer discretionary ETFs (XLK, XLY). He substitutes Bawa-Lindenberg downside semi-covariance, which uses only months when a sector's return falls below the 13-week T-bill rate. This matrix supplies both the clustering distance and the optimization. A rolling 24-month window divides Eleven Select Sector SPDR ETFs into three clusters.
A two-state Gaussian HMM then labels each month Bull or Bear. It is fitted on an expanding window using monthly SPY return and standardized VIX. The Bull objective balances the trailing 24-month mean against downside variance (γ = 1.2); in Bear, the optimizer simply minimizes downside variance. Two constraints keep the labels from changing too readily. A persistence parameter, κ = 0.30, inflates the transition matrix's diagonal. A switch also requires the other state's posterior to reach δ = 0.65. The intended payoff comes from staying with growth sectors in bull phases and retreating into XLU, XLP and XLV during 2022.
Downside covariance alone takes Plain NCO drawdown from -14.48% to -20.32%
The ablation table offers the clearest evidence about the paper's lead idea, and the author acknowledges its implications. Without a regime overlay, downside semi-covariance adds 44 bps and raises CAGR to 11.09%. Drawdown worsens from -14.48% to -20.32%. Calmar drops from 0.74 to 0.55. The author reports the deterioration directly, attributing it to downside-only minimum variance concentrating in high-upside names.
The regime layer supplies the rescue: another 232 bps, with drawdown returning to -14.22%. Section VI calls downside semi-covariance "not sufficient alone" and says sticky regime detection "restores drawdown control while retaining the return gain". Yet the ablation never pairs symmetric NCO with the sticky HMM. The return-gain part of that claim remains untested.
Several changes arrive together in that layer.
The Baseline Regime-NCO earned 11.05% using a 1D HMM on SPY returns to switch between Sharpe-max and minimum variance. Relative to that baseline, the final model changes six things at once. It introduces downside covariance and VIX as a second feature, adds κ inflation and the δ threshold, changes the bull objective, and raises the turnover penalty to 0.015 from the baselines' 0.01. The paper discloses the penalty difference. We found no arm combining symmetric covariance with the sticky HMM, nor one combining downside covariance with the naive HMM. The 232 bps therefore cannot be apportioned among persistence, VIX, the revised bull objective and the larger penalty. A trader considering stickiness for an existing rotation book gets little evidence of its separate contribution.
Sticky HMM cuts regime switches to 4 from the baseline HMM's 16
Over 65 months, the baseline HMM changed state 16 times. The sticky model changed state 4 times, a reduction of 75.0%. Monthly turnover dropped from 60.78% to 30.48%.
State changes and turnover measure different things. Even at 30.48% monthly turnover, the proposed portfolio trades more than Plain NCO at 26.99% or minimum variance at 17.13%. Only the comparison with Baseline Regime-NCO, at 60.78%, supports the paper's claim of halved turnover.
The cost table applies a flat charge per unit of turnover, ranging from 0 to 20 bps. With a 20 bps charge, the model returns 12.60% net, while its advantage over the baseline grows from 236 to 317 bps. We found no spread or market-impact model supporting this sweep. The calculation charges a flat rate against modeled turnover.
Does a nine-setting κ-δ grid with two outcomes establish a plateau?
Four switches across 65 months leave performance dependent on a handful of calls. The hyperparameter grid makes that dependence visible. For κ in {0.2, 0.3, 0.4} and δ in {0.60, 0.65, 0.70}, six settings generate exactly the same 13.41% path and four switches. The remaining three deliver 11.47% CAGR, draw down -19.22% and switch six times. Two additional regime changes cost about 194 bps annually and five points of drawdown.
Supriadi describes this as a plateau and says the selected point "sits on the edge" of it. The boundary is immediate: κ = 0.30 and δ = 0.60, one step away, puts the strategy in the bad set.
On this grid, the paper reports a CSCV probability of backtest overfitting (PBO) of 0.85. It acknowledges that the identity of the in-sample winner is "not reliably stable out-of-sample." The defence is that choosing incorrectly within the plateau carries no cost, because those six settings produce identical portfolios. With six settings tied, however, PBO alone cannot establish how often the in-sample ranking selects the weak corner. We found no count in the paper. This defence relies on the full-sample grid, with the chosen setting one step from the bad set.
The paper offers a Deflated Sharpe of 0.9976 at K = 15 trials as a counterweight. We would recalculate it first. By our arithmetic, an annual Sharpe of 0.63 across 65 months implies a t-statistic near 1.5. This is hard to reconcile with PSRs above 0.999 for every strategy, including a baseline whose Sharpe is 0.52. The authors acknowledge that the single-market, 65-month sample "limits the statistical power of the overfitting diagnostics." They do not question the PSR values themselves.
The regime layer, taken together, is the part worth keeping. Its persistence gate is the most plausible source of the improvement, though no arm separates its effect. The downside matrix is the weaker component on the evidence of its own ablation. One table could change that assessment: symmetric NCO paired with the same sticky HMM across the same 65 months. If that combination falls well short of 13.41%, downside covariance earns its place in the title.