Bursa Malaysia CAPM-GARCH tables clash with the regressions

I would use this paper's three betas and three volatility-persistence estimates, then stop. Its summary tables conflict with the regression output. We found no test showing whether the models forecast anything, and the TNB simulation appears mis-scaled.

What the paper measures

Maeenuddin, Mohd Nor and Bhatti study three Bursa Malaysia (the Malaysian stock exchange) large caps: Maybank (financials), Tenaga Nasional or TNB (utilities) and Top Glove or TGC (glove manufacturing). Their daily sample runs from January 2015 to December 2024 and contains 2,451 observations by the abstract's count. Each stock receives three treatments. An OLS CAPM regression relates daily excess returns to the FTSE Bursa Malaysia index, using the Malaysian Government Securities rate for the risk-free leg. A GARCH(1,1) estimates conditional variance. A Monte Carlo projects future prices under geometric Brownian motion using historical drift and volatility.

There is no trading rule or portfolio. The authors present the classical models as interpretable inputs to AI and ML risk systems. They argue that "integrating these statistical methodologies into an early-warning system should improve its ability to detect both micro-level asset price dynamics and macro-level contagion trends, especially in emerging economies." We found no estimated machine-learning model in the results and no contagion measure. The abstract nevertheless says "the analysis captures both micro-level asset dynamics and macro-level contagion effects." No table reports such a result. The early-warning argument remains untested.

The useful estimates

At the stock level, the fits make sense. Maybank has a beta of 0.893 and an R-squared of 0.521; TNB has 0.913 and 0.3894. TGC's beta is 1.128 (t of 10.65), with an R-squared of just 0.0443. The index therefore explains under 5% of Top Glove's daily excess-return variance. For a trader sizing a glove-stock position against a market hedge, that gap matters more than the headline beta. A beta of 1.13 suggests substantial market exposure, yet the regression leaves about 96% of variance as idiosyncratic noise. The sample includes 2020.

Volatility shocks linger across all three stocks.

Maybank's alpha1 plus beta1 is 0.976. For TGC, 0.060997 plus 0.938003 gives about 0.999. The reported TNB model is a GJR-GARCH with t errors (shape 3.004), and the paper gives its persistence as 0.999. Shocks decay slowly at those levels for both a utility and a glove maker. The authors propose regime-switching GARCH as future work; these estimates make that a worthwhile next step.

Can the volatility forecast predict anything?

The second research question asks whether GARCH can "accurately capture and predict" conditional volatility. The evidence reported is in-sample: log-likelihoods of 8282.76 for Maybank, 8102.553 for TNB and 5019.33 for TGC, alongside information criteria and residual diagnostics. Maybank also has a one-step volatility forecast of 0.00740. We did not find a holdout period, a loss function or a realised-volatility comparison used to score it.

The Monte Carlo price distributions do not use the GARCH forecasts. Tables 7, 12 and 15 draw on constant-volatility GBM with "historical mean return and standard deviation". The paper separately simulates 1,000 days of Maybank returns from GARCH, but those returns do not feed the price distributions. The three models come together in the discussion rather than in the simulations.

Tables at odds with the results

In the discussion's CAPM summary, Maybank's alpha is about +0.0032 and TNB's about +0.0048. Their regression tables instead give -0.00391 and -0.003207, both significant. The summary assigns TNB a beta of about 1.03; its estimated beta is 0.913. Maybank's R-squared of 0.521 is labelled "Low" despite being the highest of the three, while TGC's 0.0443 is labelled "Higher" despite being the lowest. A comparison table calls TNB the best CAPM fit, although Maybank has the larger R-squared.

The volatility tables have their own conflicts. The discussion cites persistence of 0.96, matching none of the three estimates. TNB's interpretation table reports alpha1 of 0.091 and beta1 of 0.913. Those sum to 1.004, at odds with the 0.999 in its GJR-GARCH summary table.

Simulated prices shift between results and summary as well. Maybank's mean is MYR 11.369 in the results and about 8.40 in the summary; its summary maximum is near 32.0, against a reported maximum of 14.280. TNB moves from RM14.95 to about 9.70.

The alphas suggest a scaling problem. A daily intercept of -0.00391 would compound to an annual abnormal return that is implausible for Maybank. We suspect the calculation subtracts an annual risk-free rate from daily returns, putting the intercept near -rf(1 - beta). TGC's positive intercept follows what that mechanism would predict for a beta above one, though it is not significant (p 0.188). The paper describes the rate as "daily interpolated". The suspected scaling error is our reading of the arithmetic.

TNB's simulation raises a similar concern. From a starting price of RM14.94, its one-year 5th to 95th percentile band spans RM14.68 to RM15.22, roughly plus or minus 1.8%. For a single equity over one year, that band is far too tight. We read it as a scaling error rather than a feature of TNB. Maybank's paths are described as "for a European call option", although the reported statistics cover stock prices and give no option value. TGC's outputs run from 0.1768 to 15.8137; the results call them returns, while the summary treats them as stock-price levels.

No local rerun

We could not test the models on our own data because we do not carry Bursa Malaysia single names. Applying the same three models to US large caps would show how US betas and GARCH persistence behave. That would address a different question from the paper's emerging-market question, so we did not substitute the US market.

What would change the verdict?

The authors acknowledge the narrow sample: "This research focuses on three Malaysian stocks; however, future studies are recommended to analyse a larger sample of stocks." The smaller problem is the number of stocks. First, the paper needs tables that agree with its regressions, a check that the risk-free leg is scaled daily, and a single rolling one-day-ahead volatility forecast scored against realised variance. Until then, its useful findings remain a 1.13 beta explaining 4.4% of Top Glove and persistence near 0.999 on two of three names.