Nonsynchronous Trading and Lead-Lag Bias in Supply Chain Equity Pairs

Abstract

International supplier-customer return predictability is unusually vulnerable to a measurement problem: two firms connected by the same economic relationship can receive different apparent lead-lag assignments solely because their exchanges close in a different order. A Korean supplier observed against a U.S. customer closes 13.5 hours earlier during U.S. daylight saving time, whereas a U.S. supplier observed against a Korean customer closes 13.5 hours later. A conventional regression that labels both observations by calendar date therefore compares different information sets. This paper constructs a close-order-adjusted lead-lag statistic for four directed supplier-to-customer relationships drawn from Altsets historical relationship data: SK Hynix to Intel, Teijin to Caterpillar, Intel to LG Electronics, and Broadcom to Samsung Electronics. The relationships deliberately reverse geographic direction while holding commercial direction constant. Daily September 2026 closes are mapped first by calendar date, then by the first economically available customer close, and finally adjusted for covariance generated by overlapping close-to-close return windows. The equal-weight mean naive slope is 0.365. Close-order alignment reduces it to 0.245. Removing the estimated overlapping-window component produces an asynchronous-adjusted mean of -0.084. All four naive coefficients are positive, while only one remains positive after both corrections. Historical relationship size, supplier revenue percentage, and customer cost percentage provide substantial variation in economic dependence, yet the residual coefficients do not sort monotonically on those measures in this four-edge experiment. The result favors a narrow interpretation: a large fraction of apparent daily international relationship predictability can be generated by return-window construction before a delayed commercial-information mechanism is needed.

Keywords: nonsynchronous trading; lead-lag; market microstructure; return synchronization; price discovery; cross-market information diffusion

Problem

Return predictability between economically linked firms has a well-established empirical foundation. Cohen and Frazzini document that returns of major customers predict subsequent supplier returns and interpret the pattern through limited investor attention to economically related firms. Their design uses observed customer relationships to identify links for which information about one company can contain valuation information about another. Shahrur, Becker, and Rosenfeld extend related reasoning internationally, examining equities across 22 developed countries and reporting that customer-industry returns lead supplier-industry returns in a manner consistent with gradual diffusion of value-relevant information. These results create an immediate econometric difficulty when the relationship crosses time zones. A positive coefficient on lagged foreign returns can represent delayed incorporation of commercial information, but it can also arise because "day t" refers to information sets ending at different UTC times. The two mechanisms have different economic content. One concerns how investors process information about connected firms. The other concerns when the dependent and independent variables were sampled.

Nonsynchronous trading has long been known to alter observed autocorrelations and cross-autocorrelations. Lo and MacKinlay develop a stochastic treatment in which nonsynchronous prices affect measured return dynamics even without the behavioral mechanism a researcher might otherwise infer from the data. In international markets, Martens and Poon show directly that close-to-close correlations can differ materially from correlations constructed from synchronized prices because exchanges operate at different times. More recent work reaches the same issue from global-factor and time-zone VAR perspectives: Asian closes can precede information contained in American closes assigned to the same calendar date, so an ordinary daily VAR or correlation model can distort the apparent direction of influence. The unresolved problem for commercially linked stocks is narrower. If the pair was selected precisely because an economic channel exists, a time-zone correction must separate genuine directional information from covariance created by the clock without discarding the relationship structure that motivated the test.

The primary hypothesis is therefore conditional rather than simply predictive. Let $s$ denote a supplier and $c$ its customer. A positive supplier-to-customer coefficient that survives actual close ordering and a mechanical overlap correction is consistent with delayed downstream incorporation of supplier information. If the positive coefficient disappears after those corrections, the original daily lead is better characterized as a sampling artifact or common-information covariance. A secondary prediction uses the direction of the Altsets dependency variables. Customer Cost %, denoted $C_{e,t}$, measures the customer's economic dependence on the supplier and should be the more relevant conditioning variable for supplier-to-customer propagation. Supplier Revenue %, denoted $S_{e,t}$, measures the supplier's dependence on the customer and is economically more natural for the reverse, customer-to-supplier direction. Relationship Size, $V_{e,t}$, measures the absolute scale of the edge. Treating all three as interchangeable versions of "relationship strength" would erase the directional mechanism the experiment is intended to test.

Timing

The September 2026 closing sequence is unusually clean. The Korea Exchange regular session ends at 15:30 Korea Standard Time, while Tokyo Stock Exchange cash trading ends at 15:30 Japan Standard Time. Nasdaq regular trading ends at 16:00 Eastern Time. Korea and Japan are UTC+9, and New York was on UTC-4 during September 2026, placing the Asian close at 06:30 UTC and the U.S. close at 20:00 UTC. The difference is 13.5 hours. This ordering implies that SK Hynix or Teijin can close before Intel or Caterpillar on the same calendar date, while Intel or Broadcom cannot close before that date's Korean customer close. A one-day lag applied mechanically to both geographic directions therefore has inconsistent information content. For an Asian supplier and U.S. customer it skips a customer close that occurs 13.5 hours after the supplier close. For a U.S. supplier and Asian customer, the next Asian trading close is correctly the first customer close available after the supplier's U.S. close.

The close itself must also be treated as date-specific in longer samples. Tokyo's cash market historically ended at 15:00, but Japan Exchange Group extended the afternoon session by 30 minutes beginning November 5, 2024. A five-year daily study that applied today's 15:30 close retrospectively would introduce a second timing error into a procedure intended to remove timing error. The correct object is therefore an exchange-calendar function $\tau_i(d)$, not a static country label. It returns the UTC close of security $i$'s exchange on trading date $d$, incorporating historical hours, daylight saving changes where applicable, and market holidays. The four-pair calculation below is confined to September 2026, so the relevant Japanese and Korean closes are both 15:30 local and no historical trading-hour splice is required for the price calculation itself.

Define the directed close-order indicator for edge $e=(s,c)$ as

For the two Asia-to-U.S. relationships, $Q_{e,d}=1$ on ordinary common trading dates. For the two U.S.-to-Korea relationships, $Q_{e,d}=0$. The naive calendar-date model is

where $n(d)$ is the customer's next local trading date after date $d$, without regard to the time at which the supplier actually closed. The information-aligned mapping instead uses

and estimates

For SK Hynix to Intel and Teijin to Caterpillar, $a_e(d)$ is generally the same calendar date while $n(d)$ is the following U.S. session. For Intel to LG Electronics and Broadcom to Samsung, the two mappings are generally identical. That asymmetry is the first measurable consequence of respecting close order.

Sample

The relationship panel contains four supplier-to-customer edges chosen to vary both economic dependence and which side of the relationship closes first. The Altsets observations are historical quarterly snapshots rather than a single current relationship flag. SK Hynix to Intel has 14 complete metric observations beginning September 2021; Teijin to Caterpillar has 21; Intel to LG Electronics has 16 beginning December 2021; Broadcom to Samsung Electronics has 10 beginning June 2023. The three dependency measures move differently across the four edges, which is useful for the identification problem. A change in Relationship Size can occur without a proportional change in either firm's dependence, while Supplier Revenue % and Customer Cost % encode opposite sides of the same directed commercial relationship.

The historical paths prevent a misleading binary strong-versus-weak classification. SK Hynix to Intel expanded from about $468 million to $1.16 billion while Intel's Customer Cost % rose from 1.22% to 2.70%; Supplier Revenue % was comparatively stable. Teijin to Caterpillar ended near the same dollar size at which it began, but Customer Cost % declined from 0.42% to 0.27%. Intel to LG Electronics became moderately larger on all three measures. Broadcom to Samsung moved in the opposite direction, with relationship size falling by roughly 55%, Supplier Revenue % falling from 1.45% to 0.31%, and Customer Cost % from 0.38% to 0.14%. These trajectories matter because a genuine supplier-to-customer return channel should be more plausibly related to the customer's cost exposure than to the supplier's sales exposure. Teijin, for example, still has a relatively high Supplier Revenue % despite a low Caterpillar Customer Cost %, making it a poor candidate for a simple undirected dependence score.

Daily adjusted closing prices are taken over the August 31 through September 2026 window required to construct September close-to-close returns. The SK Hynix and Intel series provide the Korea-to-U.S. semiconductor pair. Teijin and Caterpillar provide a Japan-to-U.S. industrial pair. LG Electronics and Samsung Electronics supply the Korean customer closes for the U.S.-originating Intel and Broadcom edges, with Broadcom's U.S. daily history providing the fourth supplier series. The return calculation uses completed local sessions only. Each log return therefore corresponds to its actual interval between consecutive local closes, rather than an assumed 24-hour interval. Weekends and holidays remain in the timing geometry because the elapsed information interval is longer even though the security itself did not trade during the closure.

Estimator

Close-order alignment by itself is insufficient. Two aligned close-to-close returns can still share many hours of the same latent information interval. Let

denote the interval underlying security $i$'s return on trading date $d$, where $d^-$ is its preceding trading date. For a source return and its aligned customer return, define the overlap

in hours. With a 13.5-hour separation, the ordinary Asia-to-U.S. aligned pair shares 10.5 hours of calendar time, while an ordinary U.S.-to-Asia aligned pair shares 13.5 hours. Weekends and holidays alter these quantities automatically. Consequently, a same-information common factor can create covariance in an aligned regression even when both stocks incorporate that factor immediately whenever their markets are open. Alignment corrects which customer close is economically available; it does not automatically make the two return intervals synchronous.

To estimate the covariance attributable to this overlap without intraday prices, consider a continuous common-information component $F_\tau$. For short increments, write the efficient returns schematically as

where $U_s$ and $U_c$ are idiosyncratic components. Under independent increments of the common component, two close-to-close returns share common-factor covariance in proportion to the length of the interval they share:

This is deliberately a parsimonious nuisance model. It is not intended as a complete return-generating process. Its role is to specify the amount of cross-product covariance that can arise mechanically from the geometry of the two return windows.

The common-overlap intensity is calibrated using the customer return ending before the supplier's current close. Let

and

After demeaning returns within each relationship, estimate

This calibration is intentionally conservative. Any covariance carried by the backward overlapping interval is assigned to the mechanical/common component, including some covariance that could in principle reflect information released before the supplier's eventual close. The resulting residual is therefore a demanding version of economic lead-lag rather than a maximal estimate of it.

The central statistic is the asynchronous-adjusted lead-lag coefficient

with

and

A positive $\widehat{\beta}^{AA}$ is the portion of the supplier-to-customer slope left after the customer return is assigned to the first close that could have incorporated the supplier close and after the overlap model removes the covariance attributable to shared return windows. A value near zero indicates that the original apparent lead can be reproduced by close ordering and shared information intervals. A negative residual can arise from sampling noise, reversal, an overaggressive mechanical subtraction, or a genuinely negative short-horizon response. With a four-edge, one-month calculation, its sign is descriptive rather than a population-level causal estimate.

Results

The naive calculation produces the pattern that a conventional relationship-momentum screen would find attractive. Every one of the four supplier-to-customer coefficients is positive. SK Hynix to Intel has a naive slope of 0.710, Broadcom to Samsung 0.466, Intel to LG Electronics 0.218, and Teijin to Caterpillar 0.066. The equal-weight mean is 0.365. The geographic split does not initially look alarming: the two Asia-to-U.S. relationships average 0.388, while the two U.S.-to-Korea relationships average 0.342. A researcher working only with date-labeled daily closes could reasonably read this as broadly positive propagation across directed commercial links, with substantial heterogeneity in magnitude. That interpretation changes when the independent variable is linked to the first customer close that actually follows it in clock time.

Close-order alignment alone reduces the equal-weight mean coefficient from 0.365 to 0.245, so approximately 67% of the original average remains before the overlapping-window correction. Almost all of that change comes from the Asia-to-U.S. side, exactly where the naive one-day convention had skipped the same-day U.S. close. The two Asia-to-U.S. coefficients average 0.388 under the naive rule and 0.148 after alignment. The U.S.-to-Korea mean remains 0.342 because the next Korean close was already the first close after the U.S. supplier close. The close-order correction therefore changes an apparently symmetric international effect into a strongly orientation-dependent one. This is the behavior expected if calendar labeling itself contributes to measured lead-lag.

The overlap correction is more consequential. SK Hynix to Intel retains a sizable 0.623 aligned slope, yet its estimated common-overlap component is 0.850, leaving an adjusted coefficient of -0.227. Broadcom to Samsung behaves similarly: a 0.466 naive and aligned coefficient is reduced to -0.147 after a 0.612 mechanical component is removed. Intel to LG Electronics is the only relationship with a positive adjusted coefficient, at 0.242. Teijin to Caterpillar remains negative after alignment and adjustment. The equal-weight adjusted mean is -0.084, compared with 0.365 naively. Expressed without treating the ratio as an inferential statistic, the positive average apparent in the calendar-day regression is eliminated under the full correction rather than merely attenuated.

The Altsets dependency histories provide a second falsification opportunity. If residual supplier-to-customer predictability is generated by a downstream commercial channel, larger Customer Cost % should at least tend to accompany larger positive residuals. That ordering is absent. SK Hynix to Intel has by far the largest 2026 Customer Cost % in the four-edge set at 2.70%, and the relationship expanded materially over the preceding five years, yet its adjusted coefficient is negative. Intel to LG Electronics has a smaller 0.68% Customer Cost % and produces the only positive adjusted coefficient. Broadcom to Samsung is economically weaker by 2026 and also has a negative residual, while Teijin to Caterpillar combines relatively high Supplier Revenue % with low Customer Cost % and a negative residual. The pattern supplies no monotonic cross-sectional support for the proposed downstream dependence mechanism in this short return sample. Supplier Revenue % also fails to organize the residuals, which is consistent with its economic role being more relevant to customer-to-supplier propagation than to the direction studied here.

Tests

The first stress test removes observations whose relevant close-to-close intervals exceed 30 hours, stripping out weekends and the longer holiday gaps that create especially large overlap intervals. This reduces the individual samples to six observations for each Asia-to-U.S. edge and ten for each U.S.-to-Korea edge, so the individual estimates become less stable, but the aggregate comparison remains informative. The equal-weight naive coefficient is 0.293. Close-order alignment cuts it to 0.068, and the asynchronous-adjusted average is -0.094. Thus the disappearance of the positive average is not generated by assigning very long weekend or holiday intervals large mechanical overlap. In this restricted set, close ordering itself performs more of the correction because the ordinary 13.5-hour exchange separation is no longer mixed with multi-day closures.

The second test removes the single largest absolute supplier return from each relationship. A one-month study can otherwise be dominated by one source-market shock, particularly when the source stock experiences an unusually large daily move. After those four observations are removed, the equal-weight naive coefficient is 0.354, the aligned coefficient is 0.224, and the adjusted coefficient is -0.005. The adjusted average therefore moves toward zero rather than recovering the original positive lead-lag. Individual relationship coefficients remain unstable, which argues against treating any one of the four adjusted estimates as a trading parameter. The cross-pair result is more stable: calendar-day regressions show positive average lead-lag, close-order alignment reduces it, and overlap subtraction leaves little positive average dependence.

These failure tests also clarify what the result is and is not identifying. The calculation does not require the proposition that international commercial information never travels slowly. Cohen and Frazzini and Shahrur, Becker, and Rosenfeld provide evidence that economically linked returns can exhibit delayed predictability in broader samples. The narrower finding is that daily international pair regressions can attribute substantial covariance to such a mechanism before first establishing that the dependent return was measured after the proposed source information and before accounting for the hours already shared by the two close-to-close windows. This distinction is particularly relevant when the link sample itself has geographic structure, because an Asia-to-U.S. edge and a U.S.-to-Asia edge are subjected to different implicit lag operators by the same calendar-day code.

Interpretation

The result changes the useful null hypothesis for international relationship research. A conventional null of $\beta=0$ asks whether one linked firm's return predicts another. The more demanding null asks whether the observed coefficient exceeds the amount that can be generated by close ordering, overlapping information windows, and broad common shocks. Those are not equivalent questions. Martens and Poon show that synchronization can materially alter international covariance estimates, while Lo and MacKinlay show more generally that nonsynchronous observation can change return cross-autocorrelations. A relationship-conditioned study adds one complication: commercial edges are selected precisely because the firms may share industries, production cycles, technology shocks, or macro exposures. Mechanical covariance is therefore not an abstract concern at the margin. It can be large for the same pairs for which economic transmission is most plausible.

The three Altsets metrics sharpen that distinction rather than merely supplying larger weights to "important" relationships. Relationship Size captures absolute commercial scale. Supplier Revenue % expresses how much of the supplier's sales base is tied to the customer. Customer Cost % expresses how much of the customer's cost base is tied to the supplier. A downstream supplier-return signal should therefore have a different economic interaction with the two percentage measures. If supplier news contains information about the customer's input costs, continuity of supply, or production economics, $C_{e,t}$ should be the more natural amplifier. If customer news forecasts the supplier because the customer represents a material revenue source, $S_{e,t}$ becomes the natural amplifier. Using a single symmetric relationship score would blur that distinction and could create an apparent dependence interaction even when the predictive direction runs opposite to the economic exposure being invoked.

The four histories also show why a contemporaneous dependency snapshot can be an inadequate conditioning variable. SK Hynix to Intel's relationship size more than doubled between the first and latest observations while Supplier Revenue % stayed near its initial level and Customer Cost % rose markedly. Broadcom to Samsung moved sharply downward on all three measures. Intel to LG Electronics strengthened more gradually. Teijin to Caterpillar's dollar relationship was roughly unchanged from its first to latest observation even as both percentage measures declined. A larger future panel could therefore estimate $\beta^{AA}_{e,t}$ against lagged $V_{e,t}$, $S_{e,t}$, and $C_{e,t}$ separately rather than ranking static edges. The present experiment supplies no basis for claiming that historical strengthening mechanically creates return predictability. Its contribution is that the dependence variables can be tested only after the return clock has been made economically coherent.

There is also a trading implication, although the four-pair calculation is not a backtest. A strategy that buys the customer after observing a foreign supplier close must define the tradable information set by timestamp. For an Asian supplier and U.S. customer, the U.S. customer has not yet reached its regular close when the supplier close becomes observable, so a next-day signal sacrifices the same-day U.S. price-discovery interval and may mistake subsequent autocorrelation for delayed transmission. For a U.S. supplier and Korean customer, the Korean cash market has already closed by the time the U.S. supplier closes, so the next Korean session is the first regular close that can incorporate the completed U.S. return. One fixed "lag one day" implementation therefore represents different strategies on the two sides of the Pacific. Any market-neutral, cross-market, or relationship-momentum test that ignores that distinction embeds a directional timing bet before portfolio construction begins.

Conclusion

International supplier-customer lead-lag is partly a clock problem. In the four directed relationships studied here, all naive calendar-day supplier-to-customer slopes are positive and their equal-weight mean is 0.365. Mapping each supplier return to the first customer close that actually follows it lowers that mean to 0.245, with the reduction concentrated in Asia-to-U.S. relationships for which the naive convention had skipped an available same-day U.S. close. Modeling covariance created by the overlapping close-to-close windows removes the remaining positive average, producing an asynchronous-adjusted mean of -0.084. Restricting the sample to ordinary one-day intervals gives -0.094, while deleting the largest supplier return in each pair gives -0.005. The precise individual coefficients are sample-sensitive, but the aggregate result is consistent across those two challenges: little positive average supplier-to-customer lead-lag survives both timing corrections.

The historical Altsets variables make the null result economically informative. The four relationships span rising, stable, and declining commercial exposure, and they differ materially across Relationship Size, Supplier Revenue %, and Customer Cost %. The surviving coefficients do not order themselves by downstream customer dependence. That weakens a simple interpretation in which stronger commercial dependence automatically produces stronger daily supplier-to-customer predictability. A broader empirical implementation could find genuine residual transmission in particular industries, states, or horizons, but the required starting point is stricter: the source return must precede the target close in real time, the shared information interval must be accounted for, and the dependency metric must correspond to the proposed direction of propagation. Without those conditions, an international relationship signal can be statistically measurable while its apparent direction is partly manufactured by the sequence of market closes.

References

Altsets supply chain dataset.

Cohen, L., and A. Frazzini. "Economic Links and Predictable Returns." Journal of Finance 63, 2008.

Lo, A. W., and A. C. MacKinlay. "An Econometric Analysis of Nonsynchronous Trading." Journal of Econometrics 45, 1990.

Martens, M., and S. H. Poon. "Returns Synchronization and Daily Correlation Dynamics Between International Stock Markets." Journal of Banking & Finance 25, 2001.

Shahrur, H., Y. L. Becker, and D. Rosenfeld. "Return Predictability along the Supply Chain: The International Evidence." Financial Analysts Journal 66, 2010.

Wu. "Estimating Contagion Mechanism in Global Equity Market with Time-Zone Effect." Financial Management, 2023.

Japan Exchange Group. Trading-hours rules and November 5, 2024 cash-market extension.

Korea Exchange. Regular market trading hours.

Nasdaq. Regular market trading hours.

Altsets. Historical relationship data extract used for the four directed supplier-customer histories in this study.

Altsets research specification. Internal brief defining the financial research objective and the role of relationship information as a research input.

Altsets quantitative research specification. Internal brief requiring integration of external research, primary evidence, Altsets relationship data, and question-specific quantitative analysis.