A trader should care about the gap between 0.9622 and 0.9211. The first is full-sample GARCH persistence for the Nepal Stock Exchange (NEPSE) index; the second comes from the 880 observations after 13 July 2022. Persistence before the break is 0.9396, estimated from 269 observations. Shrestha and Pokhrel report half-lives of 18.00 trading days for the full sample, 11.12 days before the break and 8.43 days afterward. A book sized on an 18-day decay from the full-sample GARCH therefore uses a parameter suited to neither part of the sample.
The study covers a market that rarely receives this kind of volatility analysis. Its data are daily NEPSE index closes from nepsealpha.com, spanning 30 May 2021 to 15 May 2026: 1,150 prices and 1,149 log returns. ADF rejects a unit root in returns (-9.2908, p<0.01), while the index level is not rejected (-2.3801, p=0.4174). Returns have skewness of 0.4917, excess kurtosis of 2.0621 and a Jarque-Bera statistic of 249.8829. Daily standard deviation is 0.013882, around 22% annualized.
Auto-ARIMA selects an AR(3) mean equation. The ARCH-LM statistic on its residuals is 42.2621 at lag 5 (p<0.001), giving the authors grounds to model conditional variance. They estimate GARCH(1,1) and EGARCH(1,1) by maximum likelihood with standardized Student-t innovations, whose shape is around 5.05. A GJR-GARCH(1,1) provides an additional check on asymmetry, and inference relies on White sandwich standard errors.
Forecasting uses a fixed-length rolling window. The initial estimation window contains the first 919 returns. With each new observation, the parameters are estimated again and the oldest return leaves the window. This yields 230 one-step-ahead forecasts, evaluated against absolute daily return as the realized-volatility proxy. GARCH records RMSE of 0.009552 and Mincer-Zarnowitz R² of 0.058319. EGARCH records RMSE of 0.009670 and R² of 0.045760, although its MAE is lower, at 0.007406 against 0.007521. Within the sample, GJR has the highest log-likelihood (3390.569) and the lowest AIC (-5.8861). GARCH retains the lowest BIC (-5.8475). The BIC-selected Bai-Perron procedure locates a single break on 13 July 2022.
The evidence on asymmetry remains unsettled, as the authors acknowledge. EGARCH estimates a signed-shock coefficient of -0.0379 with p=0.146 and a magnitude coefficient of 0.2611 (p<0.001). The GJR asymmetry coefficient is 0.0986 (p=0.026). Their summary reads: "the leverage effect remained model-dependent rather than conclusive." Elsewhere they report skewness of +0.4917 without tying it back to this result. The heavier tail is on the right, whereas an equity index governed by the leverage story should skew in the opposite direction.
Forecasting variance for one day
The out-of-sample target is one-day-ahead conditional variance. The AR(3) mean is estimated, and within the GARCH mean equation AR(1) (0.0985) and AR(3) (0.0726) are significant. μ (p=0.079) and AR(2) (p=0.076) are insignificant. Out-of-sample return forecasts are never scored.
The deliverable can therefore support exposure scaling toward a volatility target or a VaR budget. The analysis ends with statistical loss functions. We found no position rule, transaction costs, VaR backtest or utility calculation anywhere in the paper. That scope is legitimate for a frontier-market econometrics study. A practitioner still has to provide the whole return-generating layer, leaving any claim that the GARCH-EGARCH distinction pays beyond the paper's evidence.
Conflicting loss functions
The RMSE difference is 0.000118, about 1.2% of the level, and favors GARCH. MAE favors EGARCH.
The rankings conflict.
The authors are direct about what their metrics can establish. The paper says "the reported R² values alone do not establish forecast unbiasedness or statistical significance." It follows with: "the forecasting conclusions should be based primarily on the relative RMSE and MAE values and should not be interpreted as evidence of strong absolute predictive power." Their claim is already narrow. The remaining problem is practical: the two loss measures order the models differently on this sample, and we found no Diebold-Mariano test or other test of the loss differential. Someone selecting a production model receives little direction from the comparison. That comes close to the paper's own conclusion that the models are broadly comparable.
Another result deserves attention because it runs against intuition. The symmetric model fails the sign-bias diagnostic (2.990, p=0.0029; joint 9.319, p=0.0253). The authors answer by fitting EGARCH and then GJR. Even so, the symmetric model wins two of the three out-of-sample measures.
Only 5.8% of a noisy proxy
The higher Mincer-Zarnowitz R² is 0.058319. Absolute daily return is a noisy measure of latent volatility, making a low R² partly mechanical, as the paper notes. The reported table omits the quantities that matter for sizing: the estimated intercept, the slope and a joint test of a=0 and b=1. An R² by itself cannot distinguish a forecast with a stable positive slope from one with no slope. Those cases require very different treatment inside a volatility target. We raised the same issue in another setting, where a reported accuracy measure failed to explain the model's equity curve at all (/articles/where-finsmart-s-returns-come-from).
Where the recipe ends
The estimation procedure is specified unusually well for reproduction. It uses R 4.6.0, rugarch 1.5-5, a hybrid solver, Student-t innovations, White sandwich covariance and the exact 919/230 fixed-window protocol. Three details still required judgment.
- The sample dates disagree. The abstract shown on the journal's article page gives March 2021 to March 2026. The paper's abstract and data section instead give 30 May 2021 to 15 May 2026, with 1,149 returns retrieved 20 May 2026. We followed the dates in the body.
- Bai-Perron is said to find breakpoints "in the variance process," while the figure labels the result as a break in the return series. The regression used for the test is not pinned down. Squared returns and absolute returns will not necessarily recover 13 July 2022.
- The paper says the AR(3) mean and model set were selected before estimation. We did not find a statement limiting that selection to the first 919 observations. Parameters are re-estimated during the rolling exercise; the specification is fixed.
Can a fixed 919-day window absorb the break?
The break analysis is the paper's strongest contribution and its most obvious unfinished piece. Persistence declines after the break in both models, and both half-lives contract. GARCH moves from 11.12 days before the break to 8.43 afterward; EGARCH moves from 13.52 to 7.00. Full-sample GARCH has an 18.00-day half-life, longer than either subperiod. This is precisely the Lamoureux-Lastrapes effect cited by the paper. EGARCH's full-sample half-life of 12.00 days falls between the regimes and conceals the post-break decline.
The authors state the problem plainly: "the full-sample GARCH persistence and half-life exceed both subperiod estimates, indicating that ignoring the break overstates persistence." The abstract likewise says "a structural break was identified on 13 July 2022, after which both models exhibited lower persistence and shorter volatility half-lives." Yet the forecasting study uses a fixed 919-observation window spanning the break. It advances 230 times, once for each out-of-sample forecast, while the pre-break subsample contains 269 observations. Pre-break returns remain in the estimation window even for the final forecast.
The imbalance is large: 269 observations before the break and 880 after it. The authors themselves warn that the pre-break estimate is fragile. They make no causal attribution for the break and identify regime-switching specifications as a next step. The next paper needs a break-aware or two-state benchmark, scored on realized drawdown and turnover.
Our adaptation cannot settle the paper's claim
We could not trade the paper's asset because NEPSE is unavailable on our platform. Our implementable adaptation applies the same conditional-volatility sizing idea to SPY, QQQ and IWM. The Nepal-specific parameter estimates do not transfer.
Our automated pass created a long-only overlay that scales exposure toward a 12% annualized volatility target using a one-day Student-t GARCH or EGARCH forecast. Selection between the models uses trailing 60-forecast RMSE. A failed stability check halves exposure and reduces the estimation window to 504 returns. From 2022-01-03 to 2026-08-12, the run returned 37.30% in total, with a Sharpe of 0.83, a maximum drawdown of 11.44% and 1,150 trades. Realized volatility was 9.79% against the 12% target. The overlay carried a 0.45 beta to SPY, which explains why the loss ended at 11.44%. The paper claims only "practical implications for risk assessment," while warning against treating volatility parameters as constant through time.
Those P&L figures are provisional. The run recorded 1,150 trades without attached symbols and 1,156 daily rows across a window containing 1,162 trading days. We therefore cannot verify that the exposure reported was the exposure held. The p-value thresholds governing which model may trade were also never fixed. As a result, the traded model path is not pinned down and probably changes across runs.
The paper offers no P&L for comparison. Its own performance figures concern forecast loss: RMSE 0.009552 versus 0.009670, MAE 0.007521 versus 0.007406 and R² 0.058319 versus 0.045760, all measured over 230 forecasts. A Sharpe of 0.83 and an RMSE of 0.009552 measure different things; neither verifies nor disputes the other. Every return-producing element in our run is ours: the 12% target, the position cap, the 2bp one-way turnover charge, dynamic switching and regime de-risking.
Our protocol also departs from theirs. Model selection used squared returns in our run, while the paper uses absolute returns. We omitted the ARCH(1) through ARCH(3) candidates listed in their Table 4. Evaluation covered the full 2022-2026 period instead of their 230-day holdout. This was one automated pass, with unresolved implementation gaps, rather than a verdict on the authors' work.
Two results would change my view of the paper itself. A Diebold-Mariano test could establish whether the 0.000118 RMSE gap supports a ranking. A sizing comparison could show whether break-aware persistence beats the fixed 919-day window on realized drawdown. The former would resolve the near-tie; the latter would turn 13 July 2022 into a trading rule instead of a caveat.