Carry-Forward
Holding a metric's last known value when a fresh one isn't yet available, flagged as stale — never fabricated, never dropped. In the MEI, the rule also forbids the one shortcut that would flatter the index.
Rail: Macro · Updated: 2026-07-09
What It Is
Last observation carried forward (LOCF) is an established method for handling missing data in a time series: when a new value isn't available at the moment it's needed, the most recent previously-observed value is retained — carried forward — until a fresh observation arrives. It's routine in official statistics and time-series work, where indicators report at different cadences (daily prices alongside a quarterly or annual figure), and it keeps the series continuous so aggregation doesn't fail on null values.
LOCF has a known limitation: it assumes the underlying value hasn't changed since the last observation, so if the true value has actually moved, carrying it forward introduces bias. The standard mitigation is to pair the carried value with a stale flag — metadata showing how long it's been since a real observation — so downstream users and models can discount it appropriately. Carrying a known value forward without flagging it is what causes trouble; flagging it is what makes it honest.
It's worth distinguishing carry-forward from two neighbors. Interpolation estimates a missing value by drawing a line between a past and a future known point — which means it can't be used in real time, because it needs data that hasn't happened yet. Imputation statistically models a replacement from other correlated variables. Carry-forward is the most conservative of the three: it uses only past, actually-recorded data and invents nothing.
Why It Matters for the Machine Economy
The MEI mixes cadences by design — on-chain metrics update daily, Nvidia revenue quarterly, enterprise-adoption and connectivity figures annually with long publication lags — so carry-forward is the routine, disclosed way it handles the gap between a slow metric's releases: hold the last value, flag it stale, show the data period. That much is standard. What's specific to this platform is the rule attached to it, and it exists to close a subtle failure mode.
When a source fails or a value is missing, the MEI carries forward and never renormalizes the weights. This sounds like a technicality but it's a deliberate anti-flattering safeguard. The Legal Rail is the lowest-scoring component; if the index responded to a missing rail by dropping it and redistributing its weight across the others, a data outage on the worst-performing rail would mechanically raise the headline number. In other words, a fetch failure could make the machine economy look healthier than it is. The platform forbids exactly that: the geometric mean always runs over four components, a missing one is filled with its last good value rather than dropped, and the weights never change at runtime. Carry-forward here isn't just gap-filling — it's the mechanism that guarantees a data problem can never quietly improve the score. And the two firm boundaries around it are that the platform never fabricates a value (a metric with no live verified source becomes a disclosed gap, not an invented number) and never silently drops a metric mid-stream.
Real-World Example
A dashboard combining daily market data with an annual census figure carries the last census value forward through the year, flagged as not-yet-updated, rather than dropping the daily data or modeling fake daily population changes. The MEI does the same with, say, its annual connectivity metric between OECD releases — but adds the rule that matters: if that metric (or any component) is temporarily missing, its weight is never redistributed, so the outage can't nudge the headline up. The last good value is held, flagged, and the four-component structure stays intact.
Related Terms
- MEI (Machine Economy Index) — the index whose mixed cadences require carry-forward
- Provisional Metric — the related status for a metric not yet ready to score
- Named Gap — what the platform does instead of fabricating a missing value
- Backcasting — reconstructing genuine history where the record allows