The random walk has an advantage when 250 of the target's 252 sessions are already known at the forecast origin. I read much of this paper's ranking as the product of that smoothness and unstable early HAR estimates. The rolling results support that reading: naive wins 31 of 33 QLIKE cells. The negative finding on news sentiment is useful, though it answers a narrower question than the title implies.

One news signal across 11 large US stocks

Lis, Ślepaczuk and Sakowski test whether broad financial news helps forecast single-stock volatility. Their starting point is a 363,030-document NYT corpus. They map the 359,908 articles dated 2000 to 2022 onto trading sessions, assigning all but 11. Publication after 16:00 ET moves an article to the next session. ModernFinBERT, a pretrained three-class classifier for financial sentiment, scores the articles. Those scores produce daily aggregates:

A one-day lag and 5- and 22-day means turn these into 15 text features. Every stock receives the same market-wide signal. The 11 stocks are AAPL, ABBV, ABT, AMZN, JNJ, JPM, LLY, MA, UNH, WMT and XOM.

The benchmark called HAR uses ten regressors: five volatility lags and means, plus five log-volume terms. Matched OLS models supplement it with VIX (3 features), AAII survey sentiment (12), text (15), or all three blocks (40 predictors in total). The remaining models are Elastic Net, random forest, XGBoost, a naive last-value forecast and a stacked combination, bringing the total to 11 models.

Forecasts out of sample span mid-2002 (2015 for ABBV, late 2008 for MA) through December 2022, at horizons of 1, 5 and 22 days. Evaluation uses QLIKE. Diebold-Mariano and Clark-West tests assess significance, with Benjamini-Hochberg control across 792 tests.

Naive has the lowest mean QLIKE at every horizon. Text improves on HAR in 2 of 33 cells.

How much of the Yang-Zhang target is already known?

The target takes future trailing 252-session Yang-Zhang variance estimates, averages them, then takes the annualized square root. Those estimates are dated t+1 through t+h; the predictors stop at t-1. At h=1, this leaves 250 sessions shared between the target window and the latest input. Even at h=22, an average of 239.5 of the 252 sessions used in the target are known at the origin. These nominal horizons mostly capture the drift of a one-year window. They tell us little about next week's return variance.

The authors make the distinction explicitly. As the paper puts it, "the known historical part of these overlapping estimates makes this a highly persistent, smoothed forecast object." The conclusion says that predicting this average "is not the same task as predicting the variance generated only by the next few returns." That scope is appropriate. The abstract identifies the 252-session target and conditions its conclusions on it. The title leaves that qualification out.

Mean QLIKE favors naive; cell counts favor HAR-VIX

At h=1, 5 and 22, naive records QLIKE of 0.000286, 0.000676 and 0.003806. HAR's corresponding losses are 0.000292, 0.000711 and 0.005700. Naive's advantage is therefore 2.12%, 4.86% and 33.23%.

The 33.23% gap is concentrated in three stocks: JNJ contributes 20.41 points, JPM 10.67 and AAPL 5.91. Across the 11 long-horizon cells, HAR still wins 6.

The authors locate the source in first-fold estimation. With 337 training rows, the volatility coefficients sum to 0.613 for JNJ, 0.852 for JPM and 0.943 for AAPL. AAPL's intercept is -1.147. Later folds approach about 0.97 to 0.99. The collinear OLS fit, with ten regressors on 337 rows, fails to maintain the near-random-walk level. The paper identifies these first folds as influential.

HAR-VIX wins 17 expanding cells to naive's 12, the largest cell count. Yet its mean QLIKE is worse than HAR by 43.77%, 52.44% and 38.07%. A few large losses account for that result. In JNJ at h=1, for example, HAR-VIX loses to HAR by 0.000941 in QLIKE. Mean loss and cell counts give different rankings. Both matter.

Text's 2 wins in 33 cells disappear with long-bandwidth HAC

Only JNJ and JPM at h=22 favor HAR-Text over HAR, by 0.008653 and 0.003374 in QLIKE. In those same cells, naive's gains over HAR are larger: 0.012795 and 0.006692. As the authors observe, the text wins fit with HAR's weakness in those cells.

Newey-West bandwidth h produces 2 favorable rejections for text and 30 adverse ones. Widening the bandwidth to 252+h-1 accounts for the overlapping estimation windows; the counts drop to 0 favorable and 3 adverse. Some of the 30 adverse rejections have p-values as small as 10^-44. Only 3 retain significance with the 252+h-1 bandwidth allowing for dependence.

Under the short bandwidth h, Clark-West finds significance for text in 13 cells. This is a one-sided test on adjusted squared error. A favorable CW result alone does not establish an improvement in QLIKE, and the paper distinguishes the two.

Text fares somewhat better in the ex-post regime split. With expanding windows, HAR-Text wins 10 high-volatility cells and 2 low-volatility cells. Even so, its median relative QLIKE improvement is -8.73% in the high state and -56.28% in the low state.

Specification, tuning and the dated classifier

The estimation description reaches down to the Yang-Zhang k, block boundaries beginning at 315+2h, the purge of h, complete hyperparameter grids and seed 42.

For the finance-relevance screen, the paper uses a "predefined financial lexicon"; we did not find its term list. Section 2.1 says headline, abstract, lead paragraph and snippet are concatenated. Later, the lead paragraph alone "serves as the analyzed textual variable." Section 4.3 describes the full fold and regime tables as unavailable, although Appendix D prints them for HAR-Text.

Tuning uses the first 252 observations and stays frozen thereafter. Its criterion is MSE, whereas evaluation uses QLIKE. The first inner training set at h=22 contains 41 rows. After that, each series receives only five refits. XGB-Text chooses among 108 configurations on this sample, then posts a large loss: 0.054099 at h=1, compared with HAR's 0.000292. That loss reflects the frozen early tuning.

The stacked combination puts a ridge penalty of 1 on unscaled forecasts. Its coefficient sums under expanding windows have a median of 0.00477, leaving it close to a constant. The authors disclose that its 66 tests can relax the BH thresholds applied to the other models.

The classifier poses the deeper problem. ModernFinBERT's weights are dated July 4, 2025; their training cutoff is undocumented. The authors describe the text results as an "ex-post measurement experiment." Against LLM reference labels, agreement reaches a Cohen's κ of 0.214. Bullish recall is 0.288.

We could not rerun the study. Our news coverage begins around 2020, and our equity prices around 2010. We lack AAII history and a copy of the 2000 to 2022 NYT corpus. Substituting a newer news feed would change the information set being tested.

The practical lesson concerns the benchmark. With a heavily smoothed target, any additional channel deserves comparison against the naive forecast as well as fitted HAR. My reading of the text result would change with one evaluation: a rerun on next-h realized variance, repeating tuning at every refit.