South Korea Investor Sentiment Index

The Korea investor-sentiment suite is a monthly, top-down Baker-Wurgler sentiment composite (KRSENT) reported alongside two companion intensity transforms, the margin-loan intensity (KRMARGIN) and the foreign net-buying intensity (KRFFLOW). KRSENT is the reduced-form, top-down gauge of Baker and Wurgler (2007), the first principal component of a panel of standardized Korean investor-demand proxies rather than a bottom-up model of any single behavior. It is published as a descriptive positioning gauge and not as a forecast.

The composite is the first principal component of the selected standardized proxies.

where $s_{j,t}$ is the business-cycle-cleaned, standardized, timing-selected proxy $j$, $w_j$ its final loading, and $N = 4$. The frozen Korean panel holds turnover, the margin-loan intensity, individual net-buying intensity, and customer-deposit growth, all investor-demand or investor-behavior measures oriented so that higher readings are more bullish (Baker and Wurgler 2006; Byun and Kim 2013).

What the Korean panel cannot contain is disclosed rather than papered over. The cleanest general proxy of Baker and Wurgler, the closed-end fund discount, is dropped permanently because Korea has no institutional closed-end equity-fund market whose discount could carry the retail-sentiment signal it carries in the United States (Lee, Shleifer, and Thaler 1991). The dividend premium and the two initial-public-offering proxies are dropped at the free-data rung, since neither the payer-nonpayer valuation panel nor the issuance ingredients are retrievable from open sources without fabrication (Baker and Wurgler 2007; Chung, Kim, and Park 2017). To hold the panel at its four-proxy floor, customer-deposit growth substitutes for the initial-public-offering count, a substitution warranted by the candidate list of Baker and Wurgler and by the Korean adaptation of Byun and Kim (2013), and the resulting panel therefore carries no firm-supply-response member and is driven entirely by investor demand and investor behavior (Stambaugh, Yu, and Yuan 2012; De Long et al. 1990).

The two companions are single-line intensity transforms with no estimation. KRMARGIN is the month-end margin-loan balance as a percent of market capitalization, a leverage-sentiment reading (Byun and Kim 2013), and KRFFLOW is the monthly foreign net-purchase flow scaled by the prior month-end capitalization, the cross-border-flow reading of the foreign-investor literature (Choe, Kho, and Stulz 1999; Richards 2005; Froot, O'Connell, and Seasholes 2001). All three series describe Korean retail and cross-border positioning rather than issuing any predictive signal (Yang 2017; Ryu, Ryu, and Yang 2020).

KRSENT follows the two-stage principal-component construction of Baker and Wurgler (2006), with the loadings, the orthogonalization, and the standardization all refit on the full pinned sample at every run.

(1) Business-cycle orthogonalization. Each raw proxy is regressed on a three-regressor Korean control set and replaced by its residual, the cleaner proxy independent of the major business-cycle effects (Baker and Wurgler 2006).

The controls are industrial-production growth, retail-sales growth standing in for the three consumption categories of Baker and Wurgler, and a reference-cycle recession indicator, the three-block mapping of the five-regressor Journal of Finance control list, with employment added only if an omitted-control probe triggers (Baker and Wurgler 2007).

(2) Standardization. Each cleaned proxy is standardized to zero mean and unit variance over the pinned sample.

(3) First-stage index. The first principal component is taken over the standardized proxies and their twelve-month lags, the one-year Baker-Wurgler lag carried to monthly data (Baker and Wurgler 2007).

(4) Lead-or-lag selection. For each proxy the contemporaneous column or its twelve-month lag is retained, whichever has the higher absolute correlation with the first-stage index, the orientation-invariant reading of the Baker-Wurgler timing rule.

(5) Final index, sign, and scale. The published index is the first principal component of the selected columns, its sign flipped so the loading vector matches the maximum number of a-priori bullish signs, with ties broken toward positive turnover and then positive margin intensity, and it is rescaled to unit variance (Baker and Wurgler 2006).

The panel retains no firm-supply-response proxy once the closed-end fund discount and the initial-public-offering proxies are dropped, so the composite is built entirely from investor-demand and investor-behavior measures (Lee, Shleifer, and Thaler 1991; Chung, Kim, and Park 2017). An uncleaned companion built from the raw proxies is published alongside the orthogonalized headline, the two being near-identical, in the manner in which Baker and Wurgler report both (Baker and Wurgler 2006). The construction is a principal-component composite rather than the return-predictive partial-least-squares sentiment index of Huang et al. (2015), which is not built here for want of an obtainable implementation and is noted only as the missing robustness companion (Bouteska, Sharif, and Abedin 2024), and its measurement lineage runs through Greenwood and Shleifer (2014) and Stambaugh, Yu, and Yuan (2012).

(6) The two companion transforms. KRMARGIN is a pure ratio with no estimation, the month-end margin-loan balance $M_t$ over the same-date month-end market capitalization.

Its numerator is a market-wide margin balance covering both the KOSPI and KOSDAQ boards while the only available capitalization is the KOSPI board alone, so the level is inflated relative to a true market-wide intensity while its dynamics remain valid, and this scope mismatch is disclosed rather than repaired by fabricating a board split (Byun and Kim 2013). KRFFLOW is likewise a pure ratio, the calendar-month sum of daily KOSPI-leg foreign net purchases $F_{[t]}$ over the prior month-end capitalization, the per-market scaling of Richards (2005).

The prior-period denominator avoids the mechanical simultaneity of scaling a flow by the capitalization the flow itself moved, and the numerator is the KOSPI leg only, with the KOSDAQ foreign-flow leg excluded and the exclusion disclosed, the scope-consistent per-market fraction of the foreign-flow literature (Choe, Kho, and Stulz 1999; Froot, O'Connell, and Seasholes 2001).

Investor sentiment matters for asset prices because a class of traders acts on beliefs unmoored from fundamentals, and the risk their mispricing imposes deters the arbitrage that would correct it. In the noise-trader account of De Long et al. (1990), correlated sentiment is itself a priced risk, and Shleifer and Vishny (1997) show why rational arbitrageurs, running other people's money under short horizons, cannot fully lean against it. Baker and Wurgler (2006) build the composite this suite adapts, and the sentiment-anomaly literature documents that a broad range of cross-sectional return anomalies concentrate in high-sentiment states (Stambaugh, Yu, and Yuan 2012; Greenwood and Shleifer 2014).

KRSENT reads this literature descriptively for Korea, a market whose retail investors trade with a directness that makes a demand-side sentiment gauge legible. The Korean sentiment literature has repeatedly tied retail trading, turnover, and margin leverage to speculative episodes (Byun and Kim 2013; Yang 2017; Ryu, Ryu, and Yang 2020; Yang and Ryu 2021). The composite tracks that chronology. It rises to its sample maximum of +4.0616 standard deviations in January 2022, at the crest of the 동학개미 retail surge that carried unprecedented household participation into the market through 2020 and 2021, and it sits at its sample minimum of −0.31 in March 2020, at the pandemic-shock low from which that surge began (Choe, Kho, and Stulz 1999).

KRSENT is not a restatement of consumer confidence. Its correlation with the Bank of Korea consumer-sentiment survey is −0.078, so it measures a different object, namely the positioning revealed in the trades of investors rather than the mood reported in a household survey (Baker and Wurgler 2007). The honest use of the series is as a description of where Korean retail and cross-border positioning sit relative to their own history, and no return-predictive reading is asserted for it here, a scope the statistical testing sets out in full (Lee, Shleifer, and Thaler 1991; Park 2024).

For a markets desk the suite is a positioning and monitoring panel, not a trading signal. KRSENT summarizes how stretched Korean retail demand is relative to its own history, and the two companions decompose the reading. KRMARGIN reads the leverage channel, the margin-loan balance investors have borrowed against their positions, which the Korean evidence links to the speculative phase of the cycle (Byun and Kim 2013). KRFFLOW reads the cross-border channel, foreign net buying scaled to market size, whose persistence and contemporaneous relation to returns the foreign-flow literature has mapped in detail (Choe, Kho, and Stulz 1999; Richards 2005; Froot, O'Connell, and Seasholes 2001).

The framing the gauge licenses is historical accompaniment, not prediction. Elevated sentiment has accompanied the episodes the composite dates, and a reading near the top of its range flags positioning the market has reached before, which is surveillance information rather than a contrarian entry signal. The limits-to-arbitrage tradition is the reason for that caution, since sentiment can stay stretched and grow more stretched for long stretches, so leaning against it is itself risky (Shleifer and Vishny 1997; De Long et al. 1990; Greenwood and Shleifer 2014), and no return-predictive reading is claimed for the composite here (Stambaugh, Yu, and Yuan 2012).

The series are best read together and against the other gauges of this platform. A high KRSENT driven mainly by KRMARGIN is a leverage-led advance, while one driven by KRFFLOW is a foreign-flow-led advance, and the two have different unwinds (Baker and Wurgler 2006; Byun and Kim 2013). Because the composite is a full-sample construction whose history revises, cross-period level comparisons are approximate, and the companions, being raw ratios, are the more stable cross-period reads (Baker and Wurgler 2007; Yang and Ryu 2021).

KRSENT is a monthly Baker-Wurgler two-stage principal-component sentiment composite in standard-deviation units, and the object tested is the orthogonalized headline, its uncleaned companion carrying no separate battery and correlating with it at 0.9867. Standardization makes the level the stationary candidate, so the unit-root triplet is run on the level and the composite is not exiled to a bounded, smoothed, or filter-constructed archetype. It is unbounded in its z-scored support, so the regulated-process mechanism under which unit-root inference over-rejects does not apply (Cavaliere 2005; Cavaliere and Xu 2014). It is a static cross-sectional projection with fixed full-sample loadings rather than a two-sided time-domain filter, so its measured persistence is a property of the own dynamics of the proxies and not a smoother gain (Stock and Watson 1998). Two of its four members are unfiltered ratio levels, so it is not stationary by filter construction (Hamilton 2018). The one construction caveat carried is that turnover enters detrended by a strict trailing sixty-month moving average and customer deposits enter as a log difference, so a stationary reading partly reflects member-level construction rather than the composite alone.

Over 208 months from 2009-01 to 2026-04 the unit-root reading is ambiguous. The augmented Dickey-Fuller test fails to reject a unit root at p = 0.128 and the KPSS test rejects stationarity at p < 0.01, while the Phillips-Perron test alone rejects the unit root at p = 0.000, the sole-dissent pattern of a persistent bounded signal whose Phillips-Perron statistic is the least reliable of the three, so no clean order of integration is assigned (Said and Dickey 1984; Phillips and Perron 1988; Kwiatkowski et al. 1992; Ng and Perron 2001). The first-order autocorrelation of the level is 0.82405 with an implied half-life of about 3.58 months, a descriptive persistence summary and not an integration claim. The Ljung and Box (1978) portmanteau on the level rejects white noise at lags 12 and 24, Q = 1159.21 and Q = 1540.92 at p = 0.000 and p = 0.000, and the Bai and Perron (1998) procedure, by the Bai and Perron (2003) algorithm, finds 5 breaks in the mean at 2012-12-01, 2015-07-01, 2018-03-01, 2020-12-01, and 2023-07-01, with the Andrews (1993) sup-Wald statistic of 118.736 rejecting parameter constancy (Perron 1989). The composite has no monthly seasonal period and no posited seasonal quantity, its proxies being business-cycle-cleaned before the projection, so the seasonal-unit-root and seasonal-stationarity machinery is not applicable and is not run (Hylleberg et al. 1990; Canova and Hansen 1995; Ghysels and Osborn 2001).

The factor structure is disclosed. The first principal component explains a variance share of 0.4765 of the four-proxy panel, but a second eigenvalue of 1.0710 marginally exceeds one, so the composite is not cleanly one-factor by the Kaiser rule, and the first-stage-to-final correlation is 0.9065 (Baker and Wurgler 2006). The orthogonalization is near-inert, the headline correlating with its uncleaned companion at 0.9867 and with each business-cycle control at no more than 0.026 in absolute value, and the composite is cleanly distinct from the Bank of Korea consumer-sentiment survey, with which it correlates −0.0781 (Baker and Wurgler 2007).

The composite is substantially a margin-and-turnover object. In the leave-one-out jackknife, dropping the margin-loan proxy collapses the correlation with the full index to 0.4242 and dropping turnover to 0.8923, while dropping customer deposits or individual net-buying leaves it near one, so two of the four members are load-bearing and the proxies are not interchangeable, a named departure from the interchangeability the original composite reports. The margin and turnover loadings, 0.5470 and 0.6404, carry the index, and the deposits loading of 0.1670 is a crisis-driven, a-priori-ambiguous, non-anchor contribution that flips negative when the 2020 episode is excluded and stays negative when both crises are excluded, with no anchor member flipping sign.

The published level revises by design. The two-stage full-sample construction re-estimates its loadings, its orthogonalization, its timing selection, and its standardization at every run through the full-replace loader, so a fixed past date revises across vintages. The recursively refit index correlated with the final full-sample index at only 0.6102, 0.4734, and 0.6780 at the 2015-12, 2018-12, and 2021-12 vintages, with a maximum revision of 2.8053 standard deviations at the 2021-12 vintage, converging to 0.9998 by 2024-12, so KRSENT must not be read as a stable real-time signal at its early history (Orphanides and van Norden 2002).

The composite is published descriptively, with no return-predictive claim. On the pre-registered primary leg, the KOSDAQ−KOSPI return spread, the bias-corrected contrarian slope is negative at every horizon (−0.0024, −0.0058, −0.0269 at one, three, and twelve months) yet the one-sided bootstrap p-values (0.175, 0.2245, 0.184) do not meet the pre-registered five-percent threshold, so the index carries no contrarian-predictive claim (Stambaugh 1999). The same regression on the secondary single-index legs is statistically significant in-sample on the KOSDAQ board at every horizon (p 0.014, 0.014, 0.007) and on the KOSPI board at one and three months (p 0.021, 0.050); these secondary results are reported as computed and are expressly not the pre-registered criterion, which the composite does not satisfy.