How Long Does a Supply Chain Signal Last?

Key findings

• The supplied 561M USD HPE-Microsoft relationship provides an economically grounded path for testing information diffusion without implying one universal holding period for every Microsoft event.
• Academic evidence shows that customer information can be incorporated more quickly as supplier earnings approach, while newer peer-network research finds predictable peer effects that decay over subsequent months, supporting explicit estimation of signal decay rather than post-hoc horizon selection.

Supply Chain Signal Half: How can a quant estimate the half-life of customer-supplier information after an event and determine whether relationship strength, attention, earnings timing, and market regime change the correct holding period?

A quant can discover a real customer-supplier signal and still lose money by holding it for the wrong amount of time. Information does not diffuse across a supply chain at one universal speed. Some customer news can be incorporated into suppliers almost immediately, some can remain underappreciated for days or weeks, and scheduled supplier disclosures can accelerate the process by forcing investors to look at the customer's information more closely. For supply chain research, the useful parameter is therefore not only whether a relationship predicts anything. It is how quickly the information advantage decays after the event that activates the relationship.

The same relationship can have different half-lives around different events

Altsets shows a 561M USD relationship between HPE and Microsoft. That relationship gives a quant an economically defensible reason to test whether Microsoft information affects HPE, but it does not specify a holding period. A large Microsoft infrastructure announcement can be incorporated into HPE quickly if investors immediately understand the relevance, while a more ambiguous change in Azure or enterprise spending can take longer to translate into HPE expectations. The relationship establishes the path. The information environment determines how quickly the path is priced.

Research on the customer-supplier anomaly supports that distinction. Joshua Madsen finds that customer returns predict supplier returns shortly before the supplier's earnings announcement but not after it, and documents increased investor attention to customer information ahead of the scheduled supplier disclosure. The supplier earnings date can act like a deadline that compresses the remaining information gap. A quant who estimates one unconditional holding period across the whole calendar can therefore average together periods when customer information diffuses slowly and periods when the same information is rapidly incorporated.

Estimate decay directly instead of choosing the best holding period

A common research workflow tests one-day, five-day, ten-day, and one-month forward returns, then reports whichever horizon produces the strongest result. That creates another multiple-testing problem. A more informative design estimates the response path across several pre-specified horizons and asks how the effect decays. Local projections, distributed-lag regressions, event-time cumulative returns, and repeated cross-sectional sorts can all be used to estimate whether the signal arrives immediately, builds gradually, or reverses after overshooting.

Recent research on economically motivated peer networks provides another reason to think in terms of decay rather than one fixed window. A 2026 Journal of Financial Economics paper on dual peer effects reports that innovations in an economically motivated Peer Index are followed by higher next-month returns that gradually decay without reversal. That does not establish the same decay path for Altsets relationships, but it provides a useful benchmark for the type of hypothesis a quant can test: does information from economically connected companies diffuse into the target stock over a measurable horizon, and does the response fade because the information has been incorporated rather than because the trade reverses?

Relationship strength can define the expected speed as well as the expected magnitude

The supplied relationship data can make the test more specific. A quantified relationship such as HPE-Microsoft can be compared with structural-only Microsoft relationships in the same historical network. A researcher can test whether economically stronger customer links not only produce larger supplier responses but also shorten the time required for investors to recognize the connection. Strongly disclosed relationships may attract more analyst and investor attention, causing information to travel faster, while weaker or less obvious connections can remain underprocessed for longer.

The opposite hypothesis is also plausible. Strong relationships can create larger cash-flow consequences that take more time to estimate even when investors notice the customer news immediately. That is why the decay curve should be measured rather than assumed. Supply chain data gives the researcher a way to stratify the sample by relationship magnitude, direction, disclosure quality, and network visibility, then ask whether those properties change the speed of price discovery.

The holding period should be linked to the information process

A quant strategy should not hold a customer-news signal for a month merely because the original anomaly literature used monthly returns. If most of the measured response now occurs in the first two days, a one-month holding period can dilute the signal with unrelated risk. If the response persists for several weeks in less-followed suppliers, trading too quickly can exit before the economic information is fully incorporated. Signal half-life can therefore determine rebalance frequency, event-window design, transaction-cost tolerance, and whether the feature belongs in a continuous cross-sectional model or a short-lived event strategy.

The half-life can also vary through time as markets become better at processing relationships. More analyst coverage, better alternative data, faster information retrieval, and increasingly capable research agents can compress slow-diffusion anomalies. That possibility makes rolling or walk-forward decay estimates useful. A signal that historically lasted twenty trading days may last only three in a more efficient information environment. Treating the original horizon as permanent can make a live strategy look like it suddenly "stopped working" when the real change was faster price discovery.

Exposure freshness and return-signal decay use different clocks

An event's return response can decay while the commercial relationship remains intact. Conversely, an old relationship map can preserve familiar counterparties while its economic weights and portfolio rankings become stale. A holding-period estimate and a snapshot-refresh rule therefore need separate measurements.

A five-relationship Altsets case panel follows SK Hynix supplying Intel, Intel supplying LG Electronics, Hyundai Glovis supplying Boeing, Boeing supplying Korean Air, and PepsiCo supplying Seven & i Holdings. The historical observations are quarterly from September 2021 through June 2026. This is a small case panel, rather than a representative estimate for all supply chains. It measures similarity between network snapshots, not trading returns or the lifetime of an alpha signal.

For nonnegative edge weights, compare two snapshots with weighted Jaccard similarity: sum the smaller value for each edge across the two dates, then divide by the sum of the larger values. Keep Relationship Size, Supplier Revenue %, and Customer Cost % separate because they have different denominators. The panel also uses a bilateral ranking score, the geometric mean of Supplier Revenue % and Customer Cost %, to study how portfolio selection changes as exposures age. That constructed score measures joint dependence for this experiment; it is not a new commercial-flow estimate.

Fit an exponential summary to similarity at snapshot age $h$:

$$S_x(h)=\exp(-\lambda_x h),\qquad H_x=\frac{\log 2}{\lambda_x}.$$

The fit uses snapshot separations through 36 months. An edge can disappear and reappear, so observed similarity need not decline monotonically and the fitted curve is a summary of information loss rather than a survival law.

Source: calculations from the selected Altsets historical relationship snapshots. These are exposure-similarity half-lives, not measured return-predictability half-lives.

The ranking result explains why refresh frequency depends on the operation performed on the graph. Top-two membership overlap averages 0.526 after three months and 0.315 after six months. A modest change near a selection cutoff can change portfolio membership before the whole exposure vector changes greatly. Even surviving edges resize: at twelve months, the median absolute multiplicative error among surviving relationships is 25.6% for Relationship Size, 24.7% for Supplier Revenue %, and 20.8% for Customer Cost %.

Leave-one-relationship-out fits retain the ordering: binary overlap half-lives range from 10.4 to 25.6 months and bilateral overlap from 5.4 to 11.6 months. Alternative bilateral scores produce different exact values, which limits how literally any one refresh interval should be taken.

For a required similarity level $q$, the fitted maximum snapshot age is $h_q=-\log(q)/\lambda_x$. At 70% similarity, this panel implies roughly 9.9 months for binary topology, 3.5 months for bilateral dependence, and 3.0 months for top-two membership. Those are case-specific diagnostics for testing a refresh policy. They do not justify holding a customer-news trade for those periods. A walk-forward strategy should record both network age and time since the activating event, then evaluate its holding period and data-refresh policy separately.

Model decay as part of the feature

Supply chain data tells the quant where economically relevant information can travel. A complete strategy also needs to estimate how long that information remains underprocessed. The signal half-life should be treated as a property to measure across relationship strength, event type, attention level, and market regime, not as a holding period chosen because it maximized one historical backtest. Once the decay curve is understood, the quant can design a rebalance rule that matches the speed of the economic information instead of trading faster or slower than the relationship warrants.

The transaction-cost guide explains why the economically correct signal horizon still has to survive the costs of implementing it.

Sources

• Altsets supply chain dataset
• Madsen, Anticipated Earnings Announcements and the Customer-Supplier Anomaly
• Dual peer effects and cross-stock predictability

Related research

• When Supply Chain Signals Only Matter in Certain Market Regimes

• Does Customer Volatility Show Up in Supplier Options?

• Do Supply Chain Links Improve Covariance Models?

• Do Customer Analyst Revisions Predict Supplier Results?

Methodology

Read the methodology for this research.

Cite this research

Altsets Research. "How Long Does a Supply Chain Signal Last?." Published September 26, 2026. https://www.altsets.com/research/supply-chain-signal-half-life

Copy citationCanonical URLMethodology