Adding regime detection cut this portfolio's cumulative return to 5.44%, compared with 29.57% when the same NCO procedure used unconditional covariance. Those are the paper's own results. Its abstract nevertheless calls the outcome a success.

Inside Supriadi and Siswoyo's portfolio

Nested Clustered Optimization (NCO) groups assets by correlation, finds a minimum-variance portfolio within each cluster, then optimizes across those cluster portfolios. The clustering limits the extreme weights a noisy covariance matrix can produce. Supriadi and Siswoyo put a regime model above this process.

Each month, they refit a three-state Gaussian HMM on the trailing 24 months. The model uses four features: the LQ45 rolling mean and realized volatility, the SBN rolling mean, and the monthly CNY/IDR return. BIC selected K=3 from {2,3,4}. The final Viterbi state determines the current regime. Ranking states by their equity mean gives them the labels Bull, Sideways or Bear.

That label determines which observations enter the covariance estimate: only months within the window assigned to the current regime. If fewer than 10 qualify (N+5, with N=5), the estimate is blended toward the full-window covariance. Ledoit-Wolf shrinkage is then applied to every estimate.

There are no return forecasts. All the optimizations seek long-only minimum variance, leaving lower risk as the design's only possible payoff. A regime change triggers rebalancing, as does any weight moving more than δ from target. Trading costs are 10bp on traded weight.

The five assets are denominated in IDR: an Indonesian government bond index (SBN), the LQ45 equity index, gold, CNY/IDR and Bitcoin. Monthly data run from January 2019 to August 2025. The out-of-sample period contains 56 months beginning in January 2021, and the headline result is a maximum drawdown of 4.44%.

What happens when the regime layer is removed?

Regime-NCO versus NCO without regimes is the comparison that isolates the extra layer. The paper reports these out-of-sample results:

Every comparison goes against the regime version, including the downside measures it was meant to improve. Even plain minimum variance comes close on drawdown, at 4.75%. It beats Regime-NCO on Sharpe, -0.4408 versus -0.9522, and cumulative return, 18.35% versus 5.44%.

The abstract says the framework "successfully limits tail risk, achieving a maximum drawdown of 4.44% and demonstrating the utility of HMM-based regime detection." In the conclusion, the comparison supporting that claim is 4.44% versus 15.27% for 1/N. The authors fairly report the other side of 1/N's performance: a 76.70% return at 13.02% volatility. Buy-and-hold earned 87.98% with a 21.17% drawdown.

But 1/N allocates 20% to Bitcoin. Regime-NCO's average Bitcoin allocation ranges from 0.32% to 0.60% across the three regimes. A minimum-variance portfolio will have a smaller drawdown than an equal-weight crypto portfolio whether or not regime detection is involved. Plain NCO achieves that with half the drawdown.

The Jobson-Korkie test against plain NCO gives a p-value of 0.0790. The authors interpret this as statistical similarity between the versions. A test unable to distinguish them also cannot substantiate the abstract's claim of improvement. The point estimates all favor plain NCO: Sharpe is -0.9522 versus -0.0084, and drawdown is 4.44% versus 2.19%.

Only the comparisons with 1/N (0.0146) and buy-and-hold (0.0195) are significant at 5%. Both benchmarks have higher Sharpe ratios. The HRP results exactly duplicate plain NCO, a choice the authors disclose: HRP "is mathematically proxied by standard (regime-agnostic) NCO in this study." The abstract names a rule-based regime strategy as another benchmark. We did not find that strategy in the results table.

Labels that change almost every month

Every fit uses only the trailing window, so estimation is causal. Accuracy is a separate question. Out of sample, LQ45 returned an average -1.2294% a month in the state labelled Bull, versus +0.5563% in Sideways. Bitcoin averaged +9.3583% a month in Sideways and lost 0.6759% in Bull.

The regime changed in 55 of 56 months.

The authors acknowledge both difficulties. They describe the first as "regime drift", attributing it to the inflation and rate shock after 2020. They connect the frequent switching to HMM instability with a 24-month window. Their prescription is explicit: "enforcing transition smoothing or a state-switching penalty is essential for future practical deployments." That admission carries more weight than anything else in the paper.

For the first research question, the authors show that returns and average allocations vary across regimes. SBN averages 0.76% a month in Bull and -0.38% in Bear. LQ45's weight is 6.22% in Bear versus 16.07% in Bull. Those differences establish that the regime model changes the inputs. Whether those changes help is another question, and the plain NCO comparison says they hurt. Minimum variance is the only available action here. CNY/IDR remains the largest average holding in every regime, ranging from 34.51% to 40.52%.

An adaptive rule that trades monthly

Changing δ from 0.02 to 0.10 has only a marginal effect, according to the authors. They treat this as evidence that the adaptive framework is stable. Our reading is that the threshold barely matters when regime changes trigger trades in 55 of 56 months.

At δ=0.05, adaptive and fixed monthly rebalancing produce identical results: Sharpe of -0.9522, average monthly turnover of 27.72%, and 55 rebalances. Increasing δ to 0.10 reduces turnover to 27.04% and improves Sharpe to -0.9351. The authors acknowledge that monthly trading costs slightly depress net returns. Never rebalancing outperforms every rebalanced variant on Sharpe (-0.8118) and on the measure labelled net annualized return (8.49%).

That return label needs caution. The monthly-fixed figure is 5.44%, identical to cumulative return in the main results table. Annualized return in that table is 1.14%. One label must be wrong, which also leaves the no-rebalancing figure of 8.49% uncertain.

The data section raises two further issues. Gold is translated into IDR using CNY/IDR, although GC=F is quoted in USD and requires USD/IDR for that conversion. Costs are a flat 10bp. The authors acknowledge that this excludes slippage and market impact, while the portfolio trades more than a quarter of its value each month.

We could not reproduce the run

We could not obtain an SBN total-return series, LQ45 history or CNY/IDR spot, and therefore did not rerun the strategy. Substituting US bond and equity ETFs alongside USD crypto would create a different strategy with a different correlation structure. It would tell us nothing about these reported figures.

The paper remains useful as a candid record of what happened with a rolling 24-month, three-state HMM fitted to monthly data. Its labels changed in 55 of 56 months, and its Bull state's LQ45 mean was -1.2294% a month. I would change my view if a version using the transition penalty the authors call essential beat plain NCO's 2.19% drawdown over the same 56 months.