The random walk wins on every stock and every forecast-error metric reported here. The authors acknowledge this in their abstract. Their remaining case is that the fitted regimes can help monitor risk, a claim the paper never tests.

Inside the model

Hartanti, Nathanael, Sunaga, Nadhifah and Noviyanti fit a Gaussian hidden Markov model to three Indonesian non-primary consumer goods stocks: MNCN (media), FAST (food and beverage franchising) and ACES (specialty retail). The regime-switching machinery is familiar. A latent Markov chain moves among states, each with its own Gaussian distribution of observations. If a state persists, identifying it today should tell us something about tomorrow. The unusual choice is the input: raw daily open, high, low and close prices, with a diagonal covariance. Returns never enter the model.

The sample covers 620 trading days from January 2024 to 14 August 2026. Calibration uses the first 496 days. The remaining 124, roughly February to August 2026, form the holdout. The authors try 2 to 6 states, fitting each candidate by Baum-Welch across 376 rolling 120-day windows. Lowest average BIC selects five states for MNCN and six for FAST and ACES.

To forecast, they use Likelihood Pattern Matching, a rule taken from Nguyen (2018). The rule searches for the historical window with a model log-likelihood closest to the current window's. It then takes the next-day close change following that historical window, multiplies it by the sign of the likelihood difference, and adds the result to today's close.

Out-of-sample GHMM MAPE comes to 1.79% for MNCN, 4.65% for FAST and 6.22% for ACES. The random walk records 1.61%, 3.60% and 1.53%, respectively. The authors defend the model on different grounds: the findings, they write, "highlight the GHMM's value in characterising issuer-specific market regimes and monitoring risk, rather than improving short-horizon point forecasts." The risk-monitoring claim deserves scrutiny.

What happens to ACES?

AAE and RMSE also favour the random walk for all three names. All of these figures measure next-day price errors. The paper has no position, cost or Sharpe, leaving no measure of money made or lost.

ACES exposes the forecast rule's weakness. Persistence forecasts it most easily of the three stocks, with random-walk MAPE of 1.53%. Yet the model's error is four times larger: AAE reaches 21.11 rupiah against 5.19. The printed August rows offer an explanation. On 4 August, the model predicts 324 after a close of 350; on 5 August, it predicts 328 after 354. Every forecast through 13 August falls between 324 and 336, while actual prices range from 350 to 362.

Each displayed ACES forecast from 3 to 13 August is exactly 26 rupiah below the previous close. Under Equation 20, the observed close anchors the forecast and the signed matched change supplies the adjustment. The pattern match therefore appears to supply the same negative 26-rupiah move repeatedly. Only selected dates are printed, so the table cannot establish what happened across the full 124 days. Still, with random-walk error averaging 5.19 rupiah, that repeated 26-rupiah offset likely explains most of the model's 21.11 AAE.

Price bands with regime labels

Fitting states to price levels sorts days by price. The authors describe each state as one that "characterizes a locally stable price regime." Their table bears this out: MNCN's states occupy mostly separate bands spanning Rp 232 to 410, while FAST's stretch from Rp 158 to 880. The trend and volatility labels come later, calculated from log returns on days assigned to each state.

Mean log-return is negative in every MNCN state, ranging from -0.004% to -0.148%. A stock drifting down through five price bands would produce exactly this pattern.

The assignment of days also matters. The smoothed posterior uses observations made after the day receiving its label. A live monitor would require forward-only filtered probabilities; we did not find an evaluation of those.

FAST supplies the paper's showcase. State 2 has 8.170% volatility, occurs on 7.1% of days and carries a +0.150% trend. Together, states 2, 3 and 4 account for about 38% of observations. The authors suggest treating entry into such states as an early warning. They build no exposure rule around that warning and compare no drawdowns. Its usefulness remains a hypothesis.

FAST's six-state choice rests on 1.93 BIC points

For FAST, average BIC is 4,195.58 with six states and 4,197.51 with five, a gap of 1.93 points. ACES favours six by 5.73, with 4,008.57 against 4,014.30. MNCN alone shows a clear separation, at 20.01. Both the five-state and six-state FAST averages omit 2 of 376 windows because they failed to converge. The lower state counts use the full set, so the averages cover different sets of windows. The authors disclose the exclusions.

Six is also the edge of the search grid. For FAST and ACES, the BIC minimum therefore leaves open whether a seventh state would have won. AIC and HQC continue to fall at six for all three stocks. A six-state model has 83 free parameters in each 120-day window.

The three seeds (42, 7 and 123) establish only that one start per window attained a finite likelihood. The authors leave more seeds for future work. FAST and ACES also report an identical highest self-transition probability of 0.9698. An implementer should check that match.

We could not run the model

Our data contains no Indonesian equity prices. We therefore cannot load MNCN, FAST or ACES at all. Our history ends in 2025 as well, before the February to August 2026 test window. Replacing these stocks with US names would test another market.

An earlier HMM review showed us the same pattern: persistent states without a forecasting gain.

A forward-filtered FAST state 2 signal would change our view if it cut exposure and delivered a lower drawdown than holding the stock over the same 124 days.