Wallet cohort analysis is a foundational technique for crypto market research. The basic approach groups wallets by some shared characteristic — first activity date, first interaction with a specific protocol, holding pattern of a specific token — and tracks the subsequent behavior of the cohort across time. The technique has produced useful analytical work and is widely deployed.
The standard approach has specific pitfalls that produce misleading conclusions when not addressed carefully. The pitfalls are not always documented in published analysis. This piece documents the most consequential ones.
The wallet-equals-user assumption
The fundamental assumption underlying most wallet cohort analysis is that wallets correspond approximately one-to-one with users. The assumption is incorrect for many wallets and the divergence affects cohort conclusions significantly.
Some wallets correspond to multiple users — exchange wallets, custody service wallets, multi-signature wallets controlled by multiple parties. Cohort analysis that treats these as single-user units misclassifies the underlying user behavior.
Some users control multiple wallets — for privacy, for operational segregation, or for various other reasons. Cohort analysis that treats these multi-wallet users as multiple distinct users overstates the user count and produces misleading conclusions about user behavior at the user level.
Some wallets are operated by automated systems rather than by users at all — bots, smart contract operations, infrastructure components. Cohort analysis that treats these as user wallets pollutes the user behavior conclusions with non-user activity.
Addressing these issues requires either wallet-clustering work to identify common ownership or explicit acknowledgment that the analysis operates at the wallet level rather than the user level. The casual wallet-equals-user assumption produces conclusions that don't actually correspond to user behavior in straightforward ways.
The cohort definition sensitivity
The conclusions of cohort analysis are sensitive to specific cohort definition choices that may not be transparent in published analysis.
The definition of "first activity" matters. First on-chain activity, first interaction with a specific protocol, first deposit at an exchange — these produce different cohorts with different subsequent behavior patterns. The choice of which to use depends on the analytical question but should be transparent.
The cohort time window matters. Cohorts defined as "users who first activated in March 2024" produce different conclusions than cohorts defined as "users who first activated in Q1 2024." The window choice affects what activity gets included and what gets excluded.
The minimum activity threshold matters. Cohorts that include any wallet with any activity differ from cohorts that include only wallets with substantial activity. The threshold choice significantly affects who is in the cohort.
Sensitivity analysis across alternative cohort definitions reveals when conclusions are robust to definition choices and when they're definition-dependent. Most published cohort analysis does not include this sensitivity work.
The survivorship bias problem
Cohort analysis that examines current behavior of historical cohorts is subject to survivorship bias. The wallets currently active from a historical cohort are not representative of the original cohort. They are the survivors — the wallets that have continued to be active.
The implication is that "users from cohort X are doing Y" claims about current behavior of historical cohorts overstate the prevalence of Y in the original cohort. The non-survivors who would have been doing not-Y are no longer in the analysis.
Addressing survivorship bias requires either explicit analysis of churn alongside current behavior or transparent framing that acknowledges the survivorship limitation. Published analysis often doesn't engage with this rigorously.
The protocol-specific labeling problem
Cohort analysis often relies on labeling of wallet activity by protocol or product. The labels are typically generated by data providers or by analytical tools and may not be accurate for all wallets.
Wallet activity that's labeled as "DeFi protocol X" may include legitimate user interactions with protocol X, but may also include MEV operations, arbitrage flows, or various other activity that's mechanically different from user behavior with the protocol. The labeled activity doesn't always map cleanly to the conceptual category the label suggests.
Cohort analysis based on labeled activity inherits the label accuracy problem. Cohorts defined as "users of protocol X" may include various activity types that aren't really protocol use in the meaningful sense.
Better practice involves either auditing the labels for the specific analysis or scoping conclusions to "wallets labeled as interacting with protocol X" rather than "users of protocol X." The reframing acknowledges the methodological dependence on the labeling accuracy.
The temporal aggregation problem
Cohort analysis typically aggregates behavior across time windows. The aggregation choices affect what patterns are visible and what patterns are masked.
Daily aggregation captures different patterns than weekly or monthly aggregation. High-frequency patterns visible at daily aggregation get smoothed out at weekly aggregation. Long-term trends visible at monthly aggregation may be obscured by short-term noise at daily aggregation.
The aggregation choice should match the analytical question. Conclusions about long-term retention should use long-term aggregation. Conclusions about short-term engagement should use short-term aggregation. Mixing aggregations within an analysis produces conclusions that may not be self-consistent.
The behavior versus identity problem
Cohort analysis often draws inferences about user identity or characteristics from observed wallet behavior. The inferences require assumptions that the underlying data doesn't always support.
"Users in cohort X are sophisticated" is an inference from behavior to identity. The behavior observed may be consistent with sophisticated users, but it may also be consistent with bots running sophisticated strategies, with copytrading from sophisticated leaders, or with various other patterns that don't actually identify the wallet operators as sophisticated.
The inference from behavior to identity is often unwarranted by the available data. Better practice limits conclusions to behavior observations rather than extending them to identity claims that the data doesn't actually support.
The cross-chain coverage problem
Wallet cohort analysis typically operates within single chains because cross-chain wallet identity is hard to establish reliably. The single-chain scope produces specific limitations on what conclusions are available.
Users active across multiple chains have their activity fragmented across the per-chain analyses. Cross-chain behavior patterns that would be visible in cross-chain analysis are invisible in per-chain analysis.
The activity that has migrated from one chain to another is captured as churn in the original chain and as new entry in the destination chain. The wallet may correspond to the same user in both cases, but the cohort analysis treats them as separate phenomena.
Addressing this requires either cross-chain wallet identity infrastructure (which is itself methodologically complicated) or transparent scoping of conclusions to per-chain behavior rather than user behavior across chains.
The publication bias problem
The published cohort analyses that get attention tend to be the ones with interesting conclusions. The analyses that produce null results or small effects are less likely to be published or to receive attention. The aggregate published literature therefore overstates the prevalence and magnitude of interesting cohort effects.
For consumers of published cohort analysis, this publication bias warrants explicit consideration. Conclusions from a single analysis showing an interesting effect deserve more skepticism than the analysis alone might suggest because the publication process has selected for interesting effects.
For analysts producing cohort analysis, the responsible practice is to publish methodology and approach openly, including for analyses that produce uninteresting conclusions, rather than only publishing the analyses with attention-worthy results.
What useful cohort analysis actually looks like
The methodological pitfalls described above don't mean cohort analysis is useless. They mean that useful cohort analysis requires specific practices that distinguish it from casual application of the technique.
Useful analysis is transparent about wallet-versus-user choices and considers sensitivity to alternative assumptions where conclusions depend on them.
It documents cohort definition choices and considers sensitivity across alternative definitions.
It addresses survivorship bias explicitly when examining current behavior of historical cohorts.
It is appropriately cautious about labeling-dependent conclusions and about inference from behavior to identity.
It uses temporal aggregation that matches the analytical question rather than mixing aggregations inappropriately.
It scopes cross-chain conclusions appropriately given single-chain analytical limitations.
It is open about methodology including for results that aren't dramatically interesting.
Each of these practices adds rigor that casual cohort analysis lacks. The aggregate effect is conclusions that are credibly supported by the underlying data rather than inferences that exceed methodological foundation.
What this means for the field
The crypto market analysis field has produced enormous amounts of cohort analysis output, much of which has methodological weaknesses described above. Improving the methodological standards across the field would improve the quality of conclusions available for serious analytical work.
The improvement requires both producer responsibility (analysts working more carefully) and consumer responsibility (readers evaluating methodology more critically). Both directions are achievable but neither is currently common practice across the broader analytical community.
Methodological literacy is part of what serious engagement with this market requires. Readers who can evaluate methodology rigorously get more value from the published analytical literature than readers who treat all analysis as equivalently authoritative. The literacy is a learnable skill that pays dividends across years of engagement with the analytical work in this space.
