For a desk sizing Bitcoin to a volatility target, this paper's most useful finding is a tie. Exposure follows target vol divided by tomorrow's forecast. Lower QLIKE means a forecast is closer to realized variance, which should steady that ratio. Whether it improves profits remains open. The authors explicitly limit their evaluation to statistical accuracy, leaving VaR, hedging and trading value for future work. The results below concern forecast loss.

The field Wafi, El-Halaby and Aboul-Dahab assembled

The authors compare 27 forecasters of next-day Bitcoin volatility on a single daily series. Econometric GARCH-family models compete with LSTM deep learning models and combinations of the two. Six entries belong to the GARCH family: GARCH(1,1), iGARCH, TGARCH, GJR, APARCH and EGARCH. Three are stacked LSTMs, trained on MSE, with 64 and 32 units, 0.20 dropout and a 30-day input window. Their inputs come from nine screened predictors, with subsets used for T and TS. The remaining 18 entries are state-augmented hybrids: they feed a GARCH-type conditional variance into the LSTM alongside its other predictors. The intended division of labour gives persistence to the econometric model and nonlinear adjustment to the network.

The predictor sets expand from technical only (T) to sentiment (TS), then blockchain activity (TSB). Dates enter the balanced sample only when every series is available; nothing is forward-filled. This leaves 2,722 days, spanning 1 February 2018 to 19 July 2025. Rolling forecasts look one step ahead. Models are ranked by QLIKE, and Diebold-Mariano tests compare them with each period's winner. There are 1,583 out-of-sample days in the normal period. The crisis samples are much shorter: 102 days for COVID-19, 92 for the China mining ban and 244 for the 2022 crypto winter.

With QLIKE of 0.1934, the technical-only LSTM leads in the normal period. Hybrid_GARCH_T records 0.2062 and TGARCH 0.2209. A hybrid has the lowest QLIKE in every crisis, although a different backbone wins each time. TS supplies the information set in two crises; TSB supplies it in one.

Do calm markets justify the extra inputs?

Technical indicators alone hold their own.

Neither family-best competitor can be statistically separated from LSTM-T. TGARCH, 9th overall, records DM 1.0520 (p=0.2930); Hybrid_GARCH_T, 2nd, records DM 0.7941 (p=0.4273). The paper's authors acknowledge the tie. Plain GARCH(1,1) comes just behind TGARCH at 0.2218.

Adding data worsened the LSTM results: LSTM-TS reached 0.2332 and LSTM-TSB 0.2468. Combining models could hurt too. TGARCH was the strongest econometric entry, yet feeding it into the network gave Hybrid_TGARCH_T a loss of 0.2631. It finished below standalone TGARCH and 15th among 18 hybrids. The calm-market results give a desk no reason to buy alternative-data feeds or maintain a hybrid stack.

Three crises, changing winners

The headline losses suggest clear victories. The DM tests leave most of those victories in doubt.

The authors acknowledge the mining-ban limitation: "this episode does not offer strong evidence of statistical dominance by the hybrid model." Across six family-best comparisons, there is one rejection at 1% and one at 10%. The best standalone LSTM survives the test in every crisis. On 92 days, even an 18th-ranked model cannot be distinguished from first place. Power is weak. The authors themselves acknowledge the small daily samples in the COVID and mining-ban windows.

They describe the hybrid gains as "regime-specific," the paper's central thesis. The conclusion nevertheless says the crisis results "provide stronger evidence in favour of regime-specific hybrid modelling." A single 1% rejection across six comparisons cannot support that sentence when the winners were chosen after the event. The practical problem remains: selecting a regime-specific winner after a crisis has ended tells a desk nothing about which backbone to deploy ahead of the next one.

Each episode's winner emerged from 27 candidates, after which the other models were tested against it. We did not find a Model Confidence Set or a Reality Check-style correction. The paper first ranks all 27 models inside each window, then offers an explanation for the winner. For example, it says the choice of Hybrid_EGARCH_TSB fits the shock's crypto-specific nature.

Other details weaken the crisis crowns. The authors report "numerical instability in the rolling forecasts" for standalone APARCH, whose normal-period RMSE is 2.2598E+30, and EGARCH, whose QLIKE is 1.4943E+10. Those backbones also supply two of the three crisis winners.

The forecast target needs clarification as well. In one passage, it is the model-implied conditional standard deviation of daily returns, derived from the same return-generating process. Elsewhere, the paper describes a daily proxy "based on realized returns." Under the first interpretation, a GARCH-implied target would favour hybrids fed by GARCH by construction. Clarifying this would be my first request to the authors.

Rebuilding the technical-only comparison

We can reproduce only part of the exercise. Without the blockchain network and transaction metrics, we cannot check the TSB claims. The sentiment series is unnamed. Our crypto news coverage starts around 2020; sentiment derived from it would differ from the authors' information set, which begins in 2018. The available route is a technical-only comparison of GARCH(1,1), LSTM-T and Hybrid_GARCH_T over common dates. We then use each forecast to size a vol-targeted Bitcoin position.

The sizing exercise is ours and tests no claim made by the authors. We choose the volatility proxy and technical features ourselves. We cannot compare those features with the technical subset of the paper's nine screened predictors because the paper does not list them. Volatility-timing and portfolio optimization appear as future work in the paper; neither is run. We add turnover, exposure caps and trading costs. Evaluation of the strategy is underway, and we have no results to report.

For a sizing desk, the paper changes one thing. In a single-asset comparison covering 1,583 calm days, neither TGARCH nor the best GARCH-fed hybrid can be statistically separated from a technical-only LSTM. I would move toward hybrids if one beat LSTM-T under a multiple-comparison test during a crisis in which it had not been selected.