A trader cannot use this paper to judge whether news attention moves Indonesian apparel stocks. Its apparent directional result comes from evaluating the media coefficient at a firm size far below any company in the sample.
The sample and the result
Arief, Umyana and Yunita start from signaling theory and the Efficient Market Hypothesis, which they say holds only loosely in an emerging market such as Indonesia. Press coverage gives investors information, and investors may overreact or underreact. The authors expect coverage to move prices more on the Indonesia Stock Exchange than in developed markets because they regard it as less efficient. Mispricing would supply the edge. They choose the clothing and luxury goods sub-sector for its sensitivity to brand perception and consumer sentiment.
Of 23 listed companies in that sub-sector, Six lack complete annual reports. The remaining 17 firms provide 85 firm-years over 2020-2024. Trimming returns at |z| > 3 leaves 81 observations.
The paper measures annual returns as (Pt - Pt-1)/Pt-1 and media exposure as the natural log of each year's article count. Exposure averages 1.7056, or about 5.5 articles per company per year in the authors' translation. Firm size is log total assets (mean 28.03), with return on assets included as a control. Moderated Regression Analysis adds a media-times-size interaction to an ordinary regression. The Chow test (Prob. 0.4695 and 0.2909) favours the Common Effect Model, plain pooled OLS, over firm fixed effects.
Neither hypothesis holds up in the reported regression. Media exposure has a coefficient of -0.661 (SE 1.076, t = -0.614, p = 0.541); the interaction is 0.0227 (t = 0.587, p = 0.559). The model's F = 0.987 (p = 0.420), R² is 0.049, and adjusted R² is -0.0007. The authors reject both hypotheses.
Where does -0.661 apply?
At log assets of zero. The variables were not centred, while observed log size runs from 25.97 to 30.24. Put the interaction into the calculation and the slope at mean size (28.03) becomes -0.661 + 0.0227 × 28.03, roughly -0.03. It is about -0.07 at the smallest firm and about +0.02 at the largest.
The authors attribute the interaction null to a fairly homogeneous sample (log size SD 1.1493). As they put it, "differences in firm size in a relatively homogeneous sample are not sufficient to differentiate investor sensitivity to media coverage." The same limited spread places the uncentred -0.661 about 26 log units below the smallest observed firm.
Even so, the paper interprets its sign: "the negative direction of the coefficient indicates that increased media exposure tends to be associated with decreased returns, but this relationship is not statistically proven." That qualification matters. Across actual firm sizes, the implied slope stays near zero and changes sign. We cannot calculate standard errors for those conditional slopes because the paper gives no coefficient covariances.
Table 5 omits the interaction row; its 0.0227 is reported only in the text. The text also reverses the table's p-values for size (0.3732) and profitability (0.1430).
Wide returns, little explained
Annual returns still span -0.6667 to 1.8846 after trimming, with a standard deviation of 0.4866. The regression uses 81 observations from 17 clusters, but we found no firm clustering of standard errors. Nor does the paper include year effects, despite covering the COVID shock it invokes as context.
The media term's 1.076 SE describes a slope at log assets of zero, nowhere near an observed company. An adjusted R² of -0.0007 indicates no explanatory power for this model. It tells us little about whether coverage matters.
The authors suggest two reasons for their null. One invokes the Efficient Market Hypothesis: the market "has anticipated media information in stock prices." The other questions the measure itself: "the frequency of news coverage alone without considering sentiment and content quality is not enough to significantly influence investor" decisions. The measurement explanation is more plausible. Their anticipation account asks more of the design than it can establish.
The timing problem
The equation pairs each year's article count with that same year's return. We found no lagged specification. Price moves could therefore generate coverage just as coverage could move prices. A stock rising 188% over a year, the sample maximum, might attract articles during that year. At the mean, coverage amounts to about 5.5 articles annually (log mean 1.7056).
A significant same-year coefficient would still fall short of a tradable signal: the year's eventual article count is unknown when the year begins. Testing the anticipation account would require coverage in year t to predict the return in year t+1. The paper does not run that test. We raised a similar timing concern about same-day returns in our review of an Indonesian ESG risk study.
Replication has other obstacles. We did not find the news database used for the counts. Their minimum of 0.0000 equals ln(1), and we could not determine how firm-years with no articles were treated. The text says returns include dividends; the formula leaves them out. Requiring complete reports for all five years removes 6 of 23 firms, conditioning the sample on surviving the period with full filings.
Why we did not rebuild the count
We do not carry prices or fundamentals for these IDX names. Our US-centred news coverage cannot reconstruct annual article counts for 17 Indonesian apparel companies. Using US stocks and US news instead would examine a different market. The authors themselves warn that developed-market findings "cannot be directly applied here."
I read this paper as uninformative about media attention in Indonesian apparel and luxury goods stocks. A lagged, sentiment-signed count with year effects and firm-clustered errors would change my view if it still produced a slope near the -0.03 implied at mean size.