Coin Metrics has recalculated Ethereum’s historical Standard Flow Metrics, creating a timing complication for trading tests that rely on exchange outflows as signals.
According to an Oct. 1 notice from the digital asset data provider, Ethereum Standard Flow Metrics were recomputed from the genesis block using the most current information available as part of its Ethereum Point-in-Time release. The update encompasses all ETH Flow Metrics at both daily and hourly frequencies, with corrected historical records provided for backfilling purposes.
This introduces a practical distinction for investment research. A chart downloaded today reflects past flows viewed through the lens of subsequently acquired knowledge. Conversely, a backtest—which simulates a trading strategy using historical data—requires the specific information that was accessible at the exact moment each decision would have been made. These represent distinct datasets, even when the observations share identical dates.
The announcement does not disclose revision quantities or provide an ETH strategy comparison. Consequently, the immediate takeaway is the necessity of identifying data vintage, which refers to the specific version of the data employed in a test; measuring any actual impact on investment returns remains an open task.
Why the same historical date can tell a different story
The mechanics of Coin Metrics’ flow methodology clarify this distinction. Standard metrics incorporate all addresses currently recognized as belonging to an exchange or other monitored entity, with each address’s history tracking back to its initial nonzero balance. As new entity addresses are discovered, historical values can be restated.
Conversely, its Point-in-Time (PIT) series utilizes only those addresses known to belong to the entity during the specific historical period in question. An address begins contributing from the date of its discovery, meaning subsequent discoveries do not alter earlier PIT intervals. The provider offers daily and hourly PIT equivalents alongside its Standard exchange-flow metrics.
Attribution is the core issue here. A blockchain transfer can be retroactively attributed to an exchange once the provider successfully identifies the corresponding wallet. While a more comprehensive reconstruction helps analyze past supply dynamics using current address coverage, determining what a market participant could have realistically known requires the address data and values accessible at that earlier point in time.
Coin Metrics initially outlined the planned recomputation on Sept. 28 to highlight this distinction, anticipating completion by Sept. 30. The final notice was published on Oct. 1 at 17:04 UTC, though notice timing alone does not establish the availability date for every individual impacted value.
Furthermore, two distinct comparisons must not be conflated. Standard versus PIT evaluates different address-attribution rules, whereas comparing retained Standard history from before and after the rebuild is necessary to measure this specific revision. Because PIT represents an entirely separate attribution methodology, preserving a copy of pre-rebuild Standard values is required to maintain a specific version of the Standard product.
CryptoQuant’s documentation for ETH Exchange Flows explicitly notes that its endpoint does not support PIT accuracy. The documentation cautions that historical figures may shift as exchange wallets are discovered, incorporated, and verified via periodic clustering updates.
CryptoQuant runs automated updates weekly on Tuesdays at 00:00 UTC, pointing out that values—particularly recent observations—may fluctuate slightly. Revisions from each provider demand independent measurement and update tracking.
Consequently, for an analyst, simply recording an old query date is inadequate if historical figures are fetched anew from a mutable endpoint. While the observation dates may stay the same, the underlying data used to generate them can change.
Interpreting an outflow also calls for caution. A withdrawal simply tracks asset movement relative to attributed exchange wallets; asserting that it signifies buying activity or profitable trading demands additional supporting evidence.
Glassnode’s BTC illustration isolates data-vintage risk
Glassnode highlighted this challenge through a hypothetical backtest on March 13, 2026. The test utilized Binance’s BTC exchange balances to trigger market entry when a five-day moving average dropped below a 14-day average, and exit when the shorter average crossed above the longer one.
Spanning Jan. 1, 2024, through March 9, 2026, the backtest initiated with a $1,000 capital base and incorporated a 0.1% fee per trade. Glassnode reported rerunning the test using PIT balances while leaving the signal logic, parameters, dates, and fee structure identical, which resulted in inferior performance compared to tests using revised balances.
The key takeaway from this comparison is that keeping the trading rule constant while altering the data variant changes the outcome. A historical balance trend reconstructed with hindsight can trigger different trading decisions than one constructed using contemporaneous knowledge.
While Glassnode provided this specific BTC balance outcome, the test has not been independently replicated in this analysis. Its primary value for Ethereum research lies in the methodology: keep the rule constant and evaluate different data vintages, noting that ETH-specific signals and return impacts require dedicated experimentation.
The timeline of data availability introduces an additional constraint. Glassnode’s PIT documentation outlines two specific limitations to the assumption that past conditions can be seamlessly replayed.
First, PIT history is only available starting from the date tracking commenced for a given metric. Prior to July 2025, coverage was restricted to BTC, ETH, and a select group of tokens and metrics, before expanding across all platform metrics in July 2025. A metric introduced at that time does not gain retroactive PIT observations simply because standard historical data exists.
Second, the timestamp assigned to an observation does not necessarily reflect when a trader could have actually retrieved it. Glassnode states it has tracked relevant computed_at timestamps since September 2024—omitting the field when unavailable—and notes that API publication naturally lags behind computation.
While an unchanged historical value accounts for subsequent revisions, replaying a trading decision also requires ensuring the input data is evaluated after its actual publication time. A test that acts on data before it was accessible is effectively utilizing future information.
For Coin Metrics’ ETH data series, this underscores the necessity of documenting each metric’s initial tracking date and historical availability to customers. Similarly, Glassnode’s coverage periods and publication disclosures apply strictly to its own offerings.
The evidence needed to measure an ETH trading effect
Assessing the impact of this rebuild demands paired observations from the same data provider and metric, featuring matching exchange coverage, intervals, and dates. For evaluating revisions, this requires examining pre-rebuild Standard values alongside post-rebuild Standard history. For evaluating trading performance, it also requires an information set proven to have been available at each decision point.
The strategy rules must remain constant throughout the comparison, using identical entry and exit criteria, parameters, and evaluation windows. Availability cutoffs, execution timelines, and trading costs must be factored into the test; otherwise, altering the strategy simultaneously with the data makes it impossible to isolate the source of any performance variations.
Consequently, evaluations should differentiate between altered input values, shifted signals, modified trades, and final return variations. A data revision can alter a dataset without necessarily changing the decisions generated by a specific rule.
Ultimately, verifying the impact requires a paired ETH dataset and a fixed-rule simulation that cleanly separates data modifications from trading outcomes. While revised historical data can accurately depict supply using modern address intelligence, any assertion that exchange outflows provided a functional trading advantage demands reproducible inputs, verified publication schedules, and documented trading choices.