Random Forest wins by 0.0001 in average RMSE, too little to carry this paper's claim that machine learning beats GARCH-type models. Nageri ranks seven volatility models across nine daily exchange rate series using average error tables, without testing whether any pair of models differs. LSTM, GRU and the ensemble trail by that sliver. The paper presents the ordering as a rank rather than a thesis, even though the central ML over EGARCH conclusion depends on a handful of cells.

The comparison being made

The sample contains nine daily exchange rate series: Australia, Brazil, Canada, China, India, Japan, Switzerland, the UK and the US. Their sources are described only as "publicly available international financial databases, including central bank publications, investing platforms, and macroeconomic financial repositories." The stated period is 01 January 2015 to 08 May 2026. The paper calculates log returns as ln(P_t/P_{t-1}) times 100, then runs Jarque-Bera, ADF and Phillips-Perron, Ljung-Box, spectral peak power and dominant frequency, along with STL decomposition plots.

Seven models enter the contest. GJR-GARCH and EGARCH are the two asymmetric GARCH specifications, standard conditional variance models for the leverage effect. The five learning models are Random Forest, XGBoost, LSTM, GRU, and ENS, the paper's name for a combination of the machine learning predictions. Each country series receives an 80:20 training and testing split. MAE, RMSE and MAPE score forecasts over the held-out fifth, after which the paper averages those scores across the nine countries and ranks the models.

Random Forest finishes first on every reported metric. Its average RMSE is 0.0057, average MAE is 0.0042 and average MAPE is 11506.7780. LSTM, GRU and ENS follow at 0.0058 and 0.0043, and the paper calls them "strong". XGBoost ranks fifth with 0.0061 and 0.0045. GJR-GARCH comes sixth at 0.0507 and 0.0494. EGARCH is last, posting average RMSE of 2.2130 and MAE of 1.2519. The paper therefore argues that AI-based frameworks "should be increasingly integrated into financial risk management, exchange rate policy analysis, and macroeconomic forecasting systems."

A tradable use, if these forecasts have one, would come through volatility sizing. Better one-day-ahead variance forecasts could tighten VaR and reduce errors when pricing short-dated FX options. The paper does not make that connection. We found no Sharpe, hedging P&L, VaR coverage backtest or cost assumption anywhere in the paper. Its evidence ends with accuracy measured in RMSE units.

Which quantity is forecast?

The evaluation formulas call y_t the "actual exchange rate volatility values." The text never constructs that variable. It gives no squared-return proxy, realized volatility estimator, horizon or annualization convention. RF and EGARCH are consequently judged against a quantity whose definition we could not locate.

This omission reaches into the results. Table 1 appears to describe something other than the percent log returns specified in the methodology. Japan has a mean of 4.8091, a standard deviation of 0.1403, a minimum of 4.6042 and a maximum of 5.0853, figures that look like log levels. The paper does not identify the series or base currency behind them. Brazil is reported at -2.2120, the US at -2.2276 and Canada at 0.2792. Every Ljung-Box p-value is 0.0000, with statistics ranging from 27,166 to 29,176, a pattern produced by a level series.

Table 1's kurtosis column raises the same concern. Seven of the nine values are negative: Japan -1.2385, Brazil -1.1587, India -1.0248, US -1.0148, Australia -0.9189, Canada -0.3438 and China -0.1180. Switzerland (0.3478) and the UK (0.5063) are the only positive cases. The paper interprets the column as showing "leptokurtic tendencies, confirming the presence of volatility clustering". Platykurtic readings imply the opposite. Fat tails in daily FX returns are a settled stylized fact, yet seven of the nine cells point the other way.

If the learning models and GARCH models were fitted to different objects, their ranking would amount to a comparison of scales. The methodology leaves that possibility open.

MAPE breaks near zero

The winning MAPE for RF is 11506.7780. Taken as percentage error, that means a miss by a factor of 115. Japan's RF MAPE reaches 46,490.84. For GJR-GARCH in Japan, the reported figure is 13,324,534. EGARCH in China reaches 38,932,033.

Numbers of that size usually follow from division by actuals close to zero, exactly the problem created when a daily return or daily variance supplies the denominator. MAPE is undefined at zero and surges nearby. It says nothing useful about relative forecast accuracy here. Even so, the abstract includes it among the three metrics on which RF "achieved the best overall forecasting performance." MAPE also remains in the overall ranking table.

Removing MAPE leaves the ranking unchanged, so it carries none of the argument. Keeping it turns unstable arithmetic into headline precision.

Three cells create the EGARCH average

Three countries account for almost all of EGARCH's average RMSE of 2.2130: India at 15.6165, China at 2.9390 and Canada at 1.3100. Japan drives the 0.0507 average for GJR-GARCH, whose MAE and RMSE there are both 0.3924.

The author identifies those countries in the results. According to the paper, "the exceptionally large RMSE values for EGARCH in Canada, China, and India reveal instability and poor adaptability of traditional volatility models under changing market conditions." Its MAE discussion highlights "particularly for Japan, Canada, China, and India." The conclusion again refers to "exceptionally large MAPE values in countries such as China, India, and Canada."

The concentration is therefore visible. The paper then generalises from it. Its abstract says "traditional econometric models performed poorly," converting four country-model cells into a statement about GARCH itself.

The remaining entries in the Table 3 RMSE row tell a less dramatic story. EGARCH records 0.0084 in Australia against RF's 0.0064. The corresponding figures are Switzerland 0.0066 against 0.0046, UK 0.0086 against 0.0063, Brazil 0.0101 against 0.0070, Japan 0.0085 against 0.0065 and US 0.0090 against 0.0067. Across six of the nine countries, EGARCH runs 31 to 44 percent above Random Forest on RMSE. The gap is real and limited, far removed from the roughly 390-fold difference suggested by the averages.

Table 3 reads as a record of GARCH estimation failing in three or four country-model cells. Structural breaks may expose a genuine weakness in asymmetric GARCH. A target scaling mismatch could produce the same pattern. The reported work cannot distinguish those explanations, although that distinction decides the paper's main claim.

Another detail resists an easy explanation. LSTM, GRU and ENS have identical entries to four decimals in every one of the nine country columns across Tables 2, 3 and 4. Their shared RMSE row is 0.0064, 0.0067, 0.0036, 0.0058, 0.0048, 0.0066, 0.0047, 0.0064, 0.0068. The MAE average is 0.0043 and the MAPE average is 15814.0796. These are three separately estimated architectures, including one combination of other models' predictions, yet every column matches to four places.

Evidence the ranking needed

Diebold-Mariano tests assess pairwise equal predictive ability. A Model Confidence Set finds the group that cannot be separated from the leading model. This paper uses neither. A difference of 0.0057 versus 0.0058 across nine series is precisely where those methods matter. The same applies to the ML-versus-EGARCH claim among the six countries where EGARCH did not blow up and the gap is 31 to 44 percent rather than orders of magnitude.

Important implementation choices remain unspecified. Tuning receives one disclosure: "iterative tuning processes to improve forecasting performance and minimize overfitting," with no validation set described separately from the 20 percent test sample. The paper provides no feature set or lag structure for RF or XGBoost, no innovation distribution for the GARCH fits and no composition rule for ENS. A single fixed 80:20 split supplies the whole out-of-sample comparison, without rolling or expanding re-estimation. The base currency, quotation convention, treatment of non-trading days and sample count are all omitted. The text directs readers to "Table 6" for the weakness of GARCH. We could not find a Table 6 in the paper.

The sample also ends on 08 May 2026, although the received date is 27 January 2026. Part of the stated sample comes after submission.

We could not test these issues ourselves. The instruments are FX spot pairs spanning nine economies, while our data covers neither FX spot nor FX futures. Currency ETFs would substitute a different market. The unavailable price series blocks replication before the undefined volatility target even enters the exercise.

My view would change with a small, specific set of evidence: a stated volatility proxy, a rolling out-of-sample window, and DM statistics for the RF-versus-EGARCH pair in the six Table 3 countries where EGARCH did not blow up. If RF continues to win there with a t-statistic worth printing, the paper has a result.

As presented, the tables are more useful as a diagnostic. Four country-model cells fail catastrophically in Table 3, EGARCH in Canada, China and India plus GJR-GARCH in Japan, while the same specifications behave normally elsewhere. The paper set out to answer a different question.

We have raised the mirror-image objection before, concerning a variance premium whose reported t of 5.40 dropped below 1.2 after its legs were horizon-matched (/articles/the-correlation-rotation-premium-that-no-traded-instrument-spans). The failure mode is the same, though its direction has reversed. The mismatch inflated the premium there. Here, an undefined target may be creating the model ranking.