Chae and Kim offer a plausible return predictor for analyst-covered Korean stocks, with little evidence that a desk could trade it. Their sole portfolio result is a high-minus-low tercile spread of 0.595, with a t-statistic of 1.77. The authors treat 0.595 as significant, awarding it one star at 10%, and build their argument around the panel regressions. Those regressions leave the trading case unproven.
Measuring disagreement over target prices
TP_DISP(T) measures how far analysts' target prices differ for the same stock. The authors collect every target issued over the trailing three months, calculate their standard deviation, and divide by the mean target. TP_DISP(P), their alternative measure, uses the month-end share price as the denominator. The measures correlate at 0.93. Inclusion requires at least three targets during the window. The quarterly regressions span about 899 firms on KOSPI and KOSDAQ from January 2006 to March 2026, yielding 21,509 stock-quarters.
In the quarterly summary statistics pooled across stock-quarters, the median stock has 8 targets and dispersion of 0.098. The pooled standard deviation is 0.057. Returns are forecast for the next month, with dispersion measured only at quarter-ends: consecutive month-ends share two months of the same three-month window. The authors also check a monthly sample of 64,320 stock-months.
The proposed economics run against the familiar earnings-forecast result. Dispersion in earnings forecasts predicts lower returns, as in Diether et al. 2002 in the US and Kim and Sohn 2010 in Korea. The usual explanation invokes Miller's short-sale constraint and overpricing. Target prices incorporate long-horizon growth and sentiment, so the authors argue that their dispersion captures uncertainty about future value. Uncertain stocks are priced low and subsequently earn more.
Li et al. (2021) found this positive relation in US data and described it as an ex-ante risk premium. Chae and Kim find the same sign in Korea. Their interpretation gives information uncertainty a role alongside risk: dispersion goes beyond a simple risk proxy, with its price greatest where frictions are worst.
How much survives the portfolio sort?
The sort comes closest to something a desk could hold. Next-month excess returns increase steadily across terciles, from 0.558 for low dispersion to 1.153 for high dispersion. That ordering is clean, but the difference of 0.595, apparently percent per month, clears only the 10% significance threshold. The reported spread averages equal-weighted group means over time. The paper reports no value-weighted counterpart.
The gap between the sort's 1.77 and the panel's t=4.08 deserves the attention. With stock and quarter fixed effects in the quarterly panel, TP_DISP(T) has a coefficient of 7.753 (t=4.08). Our arithmetic using the pooled standard deviation puts the return associated with one standard deviation of dispersion at about 0.44 points a month. Include total volatility and the target-price controls, and the coefficient falls to 4.305 (t=2.15). That translates to roughly 0.25 points.
Fama-MacBeth asks the cross-sectional question that matters to a stock picker. Quarterly TP_DISP(T) coefficients range from 3.893 to 4.218, with t-stats from 1.87 to 1.98. Price-scaled TP_DISP(P) fares better: coefficients of 2.957 to 3.175 have t of 2.02 to 2.14 and clear 5%. In the monthly regressions, TP_DISP(T) coefficients range from 1.832 to 2.141, with t of 1.71 to 1.93. Monthly TP_DISP(P) delivers 1.309 to 1.535, with t of 1.78 to 2.00. Using the pooled standard deviation, our arithmetic puts one standard deviation of TP_DISP(T) at roughly 0.10 to 0.12 points a month in those monthly regressions.
The authors acknowledge the weaker estimates and emphasize that the sign holds throughout.
Disagreement over the controls gives another reason to question the panel headline. Its volatility coefficient is 70.73 (t=6.08), compared with 16.940 (t=0.67) in quarterly Fama-MacBeth. Log size has a panel coefficient of -1.851 and is insignificant in Fama-MacBeth. Stock fixed effects identify deviations from a firm's own average in a return regression. A size coefficient of that magnitude could plausibly capture a firm being temporarily expensive relative to its history.
Much of the panel's within R2 of 0.232 comes from including contemporaneous market return as a control. With market-adjusted returns instead, within R2 drops to 0.009.
The timing narrows the claim further. The authors describe the relation as stronger, "상당부분" (largely), over the last ten years, and their estimates support that reading. For 2006 to 2015, quarterly coefficients range from 3.559 to 4.183, with t of 1.13 to 1.35. For 2016 to March 2026, they rise to 8.181 to 11.570, with t up to 3.89. During the crisis windows, August 2007 to March 2009, and 2020, quarterly estimates range from 0.146 to 6.431. None is significant, with t of 0.02 to 1.00.
Small firms, heavy coverage
The mechanism rests on subgroup results estimated in separate regressions after median splits. Small firms produce 8.871 (t=3.59). For big firms, the coefficient is 4.050 (t=1.26), falling to -0.151 (t=-0.04) when volatility enters. Remove the size, volatility and turnover controls, and the big-firm estimate returns to 7.149 (t=2.24). Those controls therefore absorb the effect among big firms.
Low-turnover stocks produce coefficients of 6.864 to 8.772. High-turnover stocks range from 3.288 (t=1.13) to 5.913 (t=2.04). The estimate for low-volatility stocks is 10.295 (t=3.58), against 4.659 (t=1.75) for high-volatility stocks. IVol splits yield 8.501 against 2.023 (t=0.69); STD_ROA splits yield 6.769 against 2.720 (t=0.90). From the volatility split, the authors conclude that interpreting dispersion "단순한 위험의 대용치로" (as a simple risk proxy) is inappropriate. This conflicts with Li et al., whose US effect was stronger among high-IVol stocks.
We did not find a formal test of whether any of these paired coefficients differ. When standard errors overlap, significance in one subsample and insignificance in another establish only that one has enough power. An interaction term would settle the question. Until then, the low-volatility finding cannot separate an uncertainty discount from a risk premium.
Analyst coverage runs against the friction explanation, a tension the authors acknowledge. They expect the relation to appear among thinly covered firms, then find the reverse. High-coverage stocks yield 9.352 (t=2.65), compared with 5.115 (t=2.02) for low-coverage stocks. The high group contains 390 firms; the low group contains 829. Their explanation is measurement noise: with fewer targets, dispersion is noisier, so the estimated effect is stronger where more targets support the measure.
That explanation is fair. It still leaves coverage at odds with the friction story. Taken together, the splits suggest small, thinly traded stocks with heavy analyst coverage. The paper does not establish how often those characteristics occur together.
Still a regression short of trading
A desk needs factor-adjusted alpha for a dispersion-sorted portfolio, portfolio turnover, and returns after Korean transaction tax and spreads. We did not find these in the paper. The concentration in smaller, low-turnover stocks puts the effect where those costs are highest. The authors describe several checks without tabulating them, including SMB/HML controls and a 5-target minimum.
We could not run the test ourselves. Constructing the signal requires DataGuide's Korean target-price history: mean, standard deviation and count over trailing three-month windows back to 2006. Our coverage is US equities. Testing there would extend Li et al. while leaving this paper untested.
We would change our mind on one result: a value-weighted high-minus-low spread above 2 standard errors after Korean costs over the 2016 to 2026 sample, where the paper says the effect lives.