An equity ETF desk needs a volatility forecast that earns its keep through position sizing. Liu's review never sizes a position.
Which forecast goes in the denominator?
Start with a simple risk-targeting book: hold one equity ETF, dividing target volatility by forecast volatility to set its weight, subject to a cap. Only the forecast changes. Choosing the denominator is the trade.
Liu groups the candidates into 3 families according to their information source. Historical models come first. ARCH (Engle, 1982) and GARCH (Bollerslev, 1986) estimate persistence from past squared shocks and lagged variance. HAR combines daily, weekly and monthly realized-volatility terms to approximate long memory. Realized GARCH brings in intraday information, as does the range-based CARR, chiefly through the daily high-low range.
Options supply the next family. Black-Scholes implied volatility comes from individual quotes; model-free implied volatility combines information across strikes. Jiang and Tian (2005) found that, for S&P 500 index options, the latter contains the information in both Black-Scholes implied volatility and historical realized volatility. Machine learning supplies the remaining family: Lasso, random forests and neural networks, often attached to a HAR framework (Yang et al., 2023). The more elaborate end includes a GAT-TCN network, combining graph attention with temporal convolution. Its inputs include stock-forum sentiment scored by a fine-tuned language model (Lü et al., 2025).
Each family has a different economic rationale. Historical models depend on volatility clustering. Options reflect traders' pricing of expected variance through expiry. Models drawing on several sources may earn their place if sentiment and correlation structures contain information that prices cannot account for.
A map without reported losses
Across 26 references and two qualitative comparison tables, this narrative review supplies no data of its own. We found no loss value, test statistic or Sharpe ratio reproduced from a cited study. Its verdict is negative: "existing evidence fails to support the universal superiority of any single method across all tasks." Liu attributes the absence of a ranking to differences in proxies, loss functions, sample splits and market conditions. The review also acknowledges the limits of its evidence. First among its 4 drawbacks is domestic research's concentration on the CSI 300 Index, SSE 50 ETF options or particular market phases.
A US ETF allocator gets no answer about which denominator to use. The paper makes no claim to provide one.
Information deserves more credit
Liu earns credit for asking what actually produces a forecast improvement. The fourth listed drawback says studies "rarely" separate the contribution of new information from variable selection or nonlinear structure. Consider the GAT-TCN study. It introduces sentiment and cross-stock correlation while changing the architecture. As Liu describes the ablation, sentiment adds incremental value in one specific stock sample. This separates the sentiment input's contribution within that network. It leaves unanswered whether a linear HAR supplied with the same sentiment series would capture the gain.
The options section offers a counter-example. Wan and Tian (2023) add deep out-of-the-money option volume to HAR using SSE 50 ETF options. Prediction improves, with call volume and put volume working in different directions. The design is nested: the information changes while the linear machinery stays. Ni et al. (2008) find that non-market-maker demand for volatility predicts realized volatility in the underlying stock. Across global indices, Branco et al. (2024) found no systematic evidence of nonlinear models outperforming linear ones at every horizon.
Liu concludes that information source and evaluation structure explain outcomes across studies better than model complexity does. I give information source more weight. Yang et al. (2023) report possible gains from nonlinear fitting in some Chinese high-frequency samples. But jump components and variable selection accompany it, preventing attribution of the gain to nonlinearity alone. A nonlinear model beating HAR out-of-sample with identical inputs would change my mind.
Four conditions before a ranking
Liu calls for at least 4 standardized choices: the volatility proxy, the information cutoff, the forecast horizon, and the out-of-sample loss. Each affects a daily ETF book.
The horizon mismatch is especially awkward for implied volatility. An option-implied figure prices variance through the option's expiry; a daily rebalance requires tomorrow's variance. Without matched maturities, the comparison judges the forecast against a target it was never pricing.
The premium creates another problem. Bollerslev, Tauchen and Zhou (2009) define the variance risk premium as model-free implied variance minus realized variance. Implied variance includes this compensation alongside expectations. Treating implied volatility as a realized-volatility forecast therefore introduces bias whenever the premium differs from zero. If implied variance exceeds realized variance on average, feeding it directly into the sizing rule leaves the ETF underweight. Liu reports that Chinese studies find premium correction likely reduces forecast bias. That adjustment deserves testing.
I disagree with the text's treatment of loss functions. Liu contrasts MSE with QLIKE, saying the latter can "decrease the effect of measurement errors of proxy variables on model rankings". Patton (2011), cited in the paper, shows that both MSE and QLIKE rank forecasts consistently when the proxy is noisy but unbiased. Their weighting of under-prediction differs. That difference matters when a low forecast can send the sizing rule's exposure soaring.
The paper's sharpest sentence is also Liu's: "A lower error indicates that the forecast is closer to the proxy variable." The following sentence assigns the test of economic worth to option pricing, VaR coverage or investment decisions. Such joint tests are scarce, the third of the review's 4 drawbacks.
Our ETF experiment remains unrun
We are building the experiment and have no results to report.
Its design follows Liu's 4 conditions. Every forecaster faces the same exposure rule in one daily-rebalanced US equity ETF book. Weights use information available at the prior close. Parameters are re-estimated over rolling chronological windows, with trading costs charged whenever weights change. The comparison nests HAR, HAR plus premium-corrected implied volatility, and a nonlinear model receiving exactly the HAR-plus-implied inputs. Comparing the last pair separates the contribution of information from that of the algorithm.
We will keep three assessments separate. Forecast accuracy is measured by QLIKE and MSE against realized variance. The Model Confidence Set of Hansen et al. (2011) will identify models that are statistically indistinguishable. Risk control measures the distance between realized volatility and its target. Investment performance is assessed through net-of-cost return and Sharpe.
Those assessments can disagree. A forecast may lead on QLIKE yet lose after costs because it churns the position. This review offers no way to tell in advance which forecast will do that.