Cutting turnover roughly thirty-fold leaves gross returns near zero across Sikder's panels, the strongest evidence here against a tradable signal. The spectacular loss, LightGBM's -78.5% a year after costs on the Scope-Expansion panel, tells us more about turnover than about Dhaka. Sikder runs the hysteresis sweep, which varies how readily the rule changes positions, to distinguish excessive trading from weak forecasts. His conclusion agrees with ours: the signal has no exploitable value. The gross result supports that judgment. The -78.5% indicts the turnover.
The experiment
Sikder tests whether market regimes help forecast Dhaka Stock Exchange returns one and five days ahead. A three-state Gaussian HMM assigns regimes from daily DSEX index returns. The model uses a trailing 500-day window and is refit every 60 days. Ordering its states by emission variance produces Normal, High-Volatility and Stress. Forecasts and trades use only forward-filtered probabilities, available to a trader at the time. The proposed advantage comes from conditioning forecasts on the regime. A forecaster should respond differently in Stress; the trading rule stays flat there.
The five forecasters are a zero-return Naive benchmark, ridge on lags, linear SVR, LightGBM with 150 depth-4 trees, and a small LSTM. SVR and LightGBM receive 13 technical features: lagged returns, momentum, moving-average ratios, a volume z-score, realized volatility and RSI. Ridge gets only the five lagged returns. The LSTM gets a 10-day return sequence. Each model is run pooled and with regime included as a feature.
Walk-forward validation uses 500-day training windows, 150-day test blocks and an embargo matching the forecast horizon. Daily data cover 2018-2025 in three panels: 12 hand-picked liquid names, 6 randomly sampled names (seed 42), and 24 names added for wider sector coverage. Together they comprise 42 stocks across 17 sectors. Each side of a trade incurs 0.35% brokerage, with 15% VAT on that brokerage. There is also a 0.0125% CDBL charge and square-root slippage capped at 0.5%.
The paper reports that no model beats Naive. Every ML strategy loses most of its capital, with CAGRs ranging from -45.5% to -78.5%.
We could not rerun the study. Our data include neither DSE-listed equity histories nor the DSEX index. Substituting a US market would remove the price floor on which the paper's most original finding depends.
Can the regime improve a forecast?
The results say no. After Holm correction, adding regime as a feature produces no significant improvement over pooled estimation for any model at either horizon. All p-values exceed 0.10 on two panels; on the third, all are at or above 0.094. An expanding-window volatility-percentile detector, substituted for the HMM, gives LightGBM a DM of -0.68 (p=0.496). LightGBM's RMSE worsens at each step as volume, volatility and then regime are added, moving from 0.015951 to 0.016006. SHAP places regime 11th among 14 features.
The weak-form interpretation deserves more restraint. Against a zero forecast, LightGBM is significantly worse across all three panels: DM values are 3.97, 4.13 and 3.55. Ridge on lags cannot be distinguished from zero, with DM values of 1.25, 0.67 and -0.07. These results establish that these lagged returns and technical features failed to beat zero. We also did not find the loss differential's direction stated beside the DM formula. The text interprets positive DM as worse than Naive, so a replication must adopt that convention.
The bigger judgment concerns the universe. Inclusion in every panel requires full 2018-2025 coverage or continuous listing. Scope-Expansion candidates are drawn from 302 common-equity tickers with full coverage. Survivorship therefore enters the backtest, and the paper's limitations section does not discuss it.
Daily turnover of 15-65% is expensive enough
Average daily turnover reaches 15-65% for the ML strategies. Charging 0.35% per side, then VAT and slippage, is enough to explain CAGRs of -45.5% to -78.5%. Sikder makes that attribution himself. He then uses the hysteresis sweep to distinguish the effects of turnover from those of a weak signal. Given that test, the losses in the backtest table should not serve as evidence about regime trading. Yet the abstract uses precisely that framing when it says the strategy "loses the majority of capital on all three panels". We did not find either the regime-specific thresholds translating forecasts into positions or definitions for the ten hysteresis configurations. An implementer would have to supply both.
The sweep provides the useful test. It removes most of the cost, leaving gross return near zero. On the expansion panel, its best configuration has a Sharpe of -0.22; the Deflated Sharpe Ratio across ten trials is 0.0000. The comparison with buy-and-hold needs fixing. The paper describes -0.22 as "still short of Buy-and-Hold", although buy-and-hold's Sharpe on that panel is -0.79. By the reported Sharpe, the turnover-reduced rule comes out ahead. The comparison must therefore use a metric we could not find reported.
Buy-and-hold itself offers a weak benchmark: its CAGRs are -5.8%, -15.2% and -6.0%. A 7% risk-free rate also depresses every Sharpe.
What quiet prices conceal
Only 0.8% of days receive a Stress label, amounting to fifteen days. Those days cluster around the COVID decline and rebound rather than the intervening freeze. The paper explains why: "a mechanically halted price series shows zero realized volatility regardless of the underlying severity of the situation." Agreement between filtered and smoothed labels is just 45.2%. Using smoothed labels in a backtest would thus change the regime series traded more often than not.
The evidence for a freeze is less substantial than the headline suggests. An unnamed stock has a longest flat-price run of 374 trading days. Add a requirement that volume fall below 10% of a trailing median, and the run contracts to 4 days. The paper acknowledges that flat prices alone are an incomplete proxy for a frozen market. We did not find the length of the trailing window. As trading dries up, a declining median would keep that 4-day count down.
A 4-day low-volume spell within a 374-day flat-price run suggests suppressed price variation in a book where trading continued.
Excluding floor days yields -10.08% annualized, compared with -4.49% for buy-and-hold. The paper labels the figure an annualized return; its backtest table uses CAGR. The -4.49% corresponds to none of the three panels' buy-and-hold CAGRs. It therefore reflects another sample or another definition of return, without the paper saying which.
Three panels, or 24 stocks counted as 42?
The third column of the forecasting table is headed "Scope-Exp. (42 stocks)". The accompanying text describes it as "the 42-stock, 17-sector Scope-Expansion panel". The backtest table gives its third column the same heading. In the data section, however, the expansion panel contains 24 stocks; 42 refers only to "the combined panel". The limitations section refers to "the Random and Scope-Expansion panels' 30 tickers", consistent with 6 plus 24. If both tables' tests used 42 names, their expansion column includes the Fixed and Random stocks. It would then be a pooled re-test, at odds with the description of three non-overlapping panels.
The methods describe the bootstrap as "repeated on both panels", despite there being three. Bootstrap results subsequently appear for all three panels, leaving a discrepancy between the methods and results text. Sikder discloses a custom corporate-action adjustment for the Fixed panel, while the other 30 tickers use vendor series. He also discloses that an AI coding assistant executed the analysis. Raw data are on Mendeley; the cleaned panels and processing code are available only on request.
The author needs to settle the 24-or-42 count before this can be cited as three replications. For surviving DSE names over 2018-2025, the conclusion that holds is narrower than the abstract's. These features and this HMM show no forecast edge. Reducing turnover leaves no gross trading edge either. The absence of Stress labels during the freeze is a finding the paper does establish, visibly illustrated in its regime chart (Figure 1). Whether that freeze involved real execution risk remains untested. The volume-aware count covers one stock's longest episode. Reporting the same count for all 42 names, together with the trailing window, would settle that question.