The monthly volatility forecast is the paper's credible result. Salisu, Ogbonna, Gupta and Bouri also report an equity portfolio with a Sharpe ratio of 17.31, a figure that deserves far more suspicion. Their low-frequency predictor spans 164-year and beats realized volatility and a standard macro factor within the same monthly forecasting model. Daily performance is weaker. At the short daily horizon, there is nothing. The monthly evidence deserves the attention.

We ran no test of any of this. The PCI is absent from our data, and reconstructing it requires the digitized nineteenth-century New York Times archive along with article-level classifications we do not hold. Nothing below is a replication. The paper forecasts volatility for the S&P 500 and the DJIA, neither of which can be held directly. A deployable strategy would therefore trade SPY or DIA instead. The broad-equity volatility mechanism could carry across to those tracking ETFs, though the missing PCI input remains unresolved.

The index beneath the forecast

The predictor comes from Jamilov (2025). His Price Conflict Index runs quarterly from 1860:Q1 to 2023:Q4 and measures disagreement over prices and wages in newspaper coverage. Jamilov sent every article in the digitized New York Times archive through a ChatGPT 4o instance. Under a fixed prompt, the classifier decided whether each excerpt mentioned conflict or disagreement among stakeholders over prices or wages in the United States. Each article received a binary flag. The resulting quarterly series has mean 2.04, standard deviation 0.88, skewness 0.73 and kurtosis 3.82.

Rowthorn's conflict theory of inflation supplies the economic argument. Jamilov finds that positive changes in PCI precede higher inflation, weaker output growth, higher unemployment and lower stock returns. Salisu and coauthors extend that sequence to volatility: if PCI contains leading bad news about the economy, the leverage effect could carry it into equity volatility. That channel gets headline billing. A footnote adds elevated risk premia and macroeconomic uncertainty, without testing either. The paper claims no return predictability from PCI. Its allocation exercise produces higher returns through volatility timing.

The model is GARCH-MIDAS in the Engle, Ghysels and Sohn form. Mixed data sampling lets quarterly PCI drive the long-run variance component through a one-parameter beta weighting function. The monthly version uses 36 MIDAS months, while the daily version uses 252 MIDAS days. Monthly S&P 500 log returns cover 1860:01 to 2023:12. Daily DJIA log returns extend from 4 May 1885 to 31 December 2023. Both series come from Finaeon's Global Financial Database.

Evaluation relies on a single 75:25 estimation and testing split. The modified Diebold-Mariano test of Harvey, Leybourne and Newbold is reported at h = 20, 60 and 120. One benchmark is GARCH-MIDAS with realized volatility. The other uses GARCH-MIDAS with the first principal component of output growth, inflation, unemployment and the short rate. Availability of the macro data limits that comparison to 1890 onward.

Where do forecast gains hold?

Against the realized-volatility benchmark, monthly DM* statistics are -4.009, -2.054 and -4.096. The short daily horizon produces -1.000, which the authors plainly describe as statistically indistinguishable from the benchmark. Daily results improve to -2.997 and -3.607 at h = 60 and 120. Against the macro principal component, monthly statistics are -1.9559, -7.0908 and -7.7335. Their daily counterparts are -4.888, -3.245 and -3.224.

GARCH-MIDAS variants have already been applied to many low-frequency behavioral, financial and macroeconomic predictors. The novelty lies in having a predictor available in 1860 that continues to contain information in 2023. The authors accordingly describe their work as the first equity volatility forecast using PCI over the longest available sample. The contribution to volatility forecasting is genuine. Against the macro factor, the monthly DM* sequence of -1.9559, -7.0908 and -7.7335 is hard to dismiss.

The allocation contribution has a much thinner base. Its comparison set contains two univariate specifications. We did not find a HAR benchmark or a joint model containing both PCI and the macro factor. The paper concludes that PCI limits "the incremental forecasting contribution of macro predictors". A pairwise comparison of single-predictor models cannot establish that the macro factor adds nothing once PCI is present. A nested specification could answer that question.

The appendix also contains evidence that strains the proposed mechanism. Of the six disaggregated conflict indexes, Energy performs significantly worse than plain realized volatility at every monthly horizon. Its DM* statistics are +4.468, +3.987 and +3.123, all at 1%. Interest Rates begins at +2.691 for h = 20, then turns negative over the longer horizons. A footnote reports gains "in many cases, though not as strong as the PCI", naming wages, goods and services, real estate and tariffs. At monthly frequency, Energy news about price conflict actively worsens the forecast. The aggregate index therefore appears to capture something broader than price-conflict content uniformly shared by its components.

The 0.0146 problem

Table 4 breaks the link between the forecast and a tradable portfolio. At the monthly frequency, the PCI portfolio returns 2.3331 versus 1.7338 for the benchmark, an increase of about 35%. Reported volatility is 0.0146 against 2.5906, lower by a factor of 177. Because the paper defines Sharpe as excess return divided by the square root of portfolio variance, the denominator directly produces 17.31.

A monthly series earning 2.33 with a standard deviation of 0.0146 does not resemble a book holding US equities.

The paper provides no diagnostics for the weight path. Equations 4 and 5 include no turnover or transaction-cost term. Even so, the authors carry the utility claim into their summing-up. The abstract says PCI "provides higher utility gains". The conclusion again cites higher Sharpe ratios and lower portfolio volatility, without addressing the 0.0146. Their only stated limitation concerns asset scope: the study covers the aggregate US market while excluding industries and bonds.

The daily panel is more believable and less striking. Under the leverage-6 setting, PCI volatility is 3.7033 versus 7.0732, with Sharpe 0.0098 against 0.0076. Under leverage-8, volatility is 6.5862 versus 12.5778, and Sharpe is 0.0080 against 0.0062.

Its accompanying text contradicts itself across consecutive sentences. The first says the PCI Sharpe "is significantly lower than those obtained from the conventional GARCH-MIDAS-RV model across both utility settings". The next says the specification "yields higher Sharpe ratios that substantially exceed the benchmark values". The table supports the second statement. The authors qualify the cross-frequency comparison by writing that "a one-to-one comparison is perhaps not completely appropriate" before judging the monthly mix to have stronger utility gains. Their qualification applies only to the monthly-versus-daily comparison.

Live implementation

The historical series is free, and the paper directs readers to Jamilov's site for the data. Updating it requires a licensed New York Times article feed, plus a classifier that reproduces the prompt. A live implementation must also choose a publication lag, an issue the finished archive lets the paper avoid. The classifier and prompt were constructed once across the whole corpus.

PCI updates four times a year. The paper distributes the quarterly long-run component over 252 days, "rolling back across days without retaining the quarterly identifiers". Four updates a year stretched over 252 trading days leave a slow-moving input for a volatility target.

We have previously argued that volatility model rankings can shift with the loss function, in our note on EGARCH versus plain GARCH on NEPSE. This paper does not identify the loss function used for DM*.

A HAR benchmark and a nested PCI-plus-macro specification would make the forecasting evidence more persuasive. For the economic evidence, Table 4 needs the realized weight path, the printed portfolio return series and explicit costs. Until those appear, the defensible claim remains narrow. A newspaper measure of price disagreement spanning 164 years improves monthly equity variance forecasts over the final 25% of one fixed 75:25 split. Its utility exercise includes no costs, no turnover and no weight path.