How Moonboard turns raw data into comparable numbers

Moonboard scans a universe of several hundred crypto assets and reduces each one to a handful of numbers. The difficulty is not the arithmetic. It is that the raw inputs are not comparable: a market capitalisation is measured in billions, a sentiment classification in points from 0 to 100, a social dominance figure in fractions of a percent, and an engagement count spans five orders of magnitude across the universe.

Everything below exists to make those quantities addable. Where a step involves a judgement call — a weight, a cut-off, a cap — the judgement is stated rather than hidden, because a weight you cannot see is a weight you cannot argue with.

1. Input data

Every metric is computed from a single observation per asset per refresh, containing these fields:

2. Cleaning and missing data

Assets arrive with incomplete records. A token listed yesterday has price and volume but no meaningful social history; a thinly covered asset may have social volume but no sentiment classification. The processing rules are deliberately conservative:

• A zero market cap voids the record. Every metric that divides by market capitalisation returns a neutral value rather than an infinity. A missing denominator is not a score of zero — it is an absence, and it is treated as one.
• Absent social fields are read as zero, not as unknown. This is a real limitation. An asset nobody discusses and an asset whose social data failed to arrive produce the same input, and the metrics cannot distinguish them. Low social scores on obscure assets should therefore be read as "no evidence of activity", never as "evidence of no activity".
• Non-finite intermediate values are dropped from the cross-section before ranking, so a single corrupt record cannot shift everyone else's percentile.

3. Normalisation

Three different techniques are used, and which one applies determines how the resulting number should be read.

Log damping

Turnover, engagement and volume are heavy-tailed: the top asset can carry a thousand times the activity of the median. Feeding those raw into a weighted sum lets a single outlier dominate every other term. Taking log₁₀(x + 1) compresses the tail while preserving order, and the + 1 keeps the function defined and non-negative at zero.

Percentile ranking

Four of the seven metrics end in a cross-sectional percentile: the share of the scanned universe scoring strictly below this asset, times 100.

This is the most consequential design decision on the platform, and it has a direct consequence: these values are relative, and they move when the universe moves. An asset can gain twenty percentile points without any change to its own data, purely because other assets cooled off. Percentile metrics answer "how does this compare right now", never "how much of this is there".

Z-scores

One metric standardises instead of ranking. Each input is expressed as distance from the universe mean in standard deviations:

Unlike a percentile, a z-score preserves magnitude — it distinguishes an asset that is slightly ahead of the pack from one that is far ahead. It also assumes a distribution where a mean and a standard deviation are meaningful descriptions, which for heavy-tailed crypto data is an approximation, not a fact. See what a z-score is and how COR applies it.

4. Weighting

Weights across the platform are fixed constants chosen for interpretability, not fitted to historical returns. Nothing here is the output of an optimisation over past performance.

That choice cuts both ways. It means the weights cannot be overfitted to a particular market regime, and it also means they are not claimed to be optimal. They encode a view about what matters — that market dynamics deserve more room than social buzz, that community standing is worth roughly as much as social momentum — and a different defensible view would produce different numbers.

5. Reading the three scale types

6. Time windows

Every social and volume field is a trailing 24-hour aggregate; volatility is realised over a recent window; market cap and price are point-in-time. All metrics are therefore snapshot metrics. They describe the state of an asset at one moment relative to its peers at that same moment.

No metric on the platform contains a lagged term or compares an asset to its own past. Trend information comes from plotting the snapshots over time, which the application does in its historical view — not from the scores themselves.

7. Outliers and caps

Several component scores are capped, for instance the volume contribution to the Moonboard Score, which reaches its maximum once turnover passes a threshold. Caps stop a single extreme input from consuming an entire score, but they also destroy information above the cap: two assets on opposite sides of an extreme become indistinguishable in that component. Where a cap binds in practice, the individual metric pages say so.

8. Data quality and what it does to the output

Scores are recomputed as new observations arrive, which is why a value can change between two page loads on moonboard.ai. Three effects follow:

• Recency. A score reflects the most recent successful collection, not the current instant.
• Coverage. Assets with sparse social data score low on social metrics regardless of their actual community.
• Composition. Every percentile metric shifts when assets enter or leave the scanned universe.

Data sources and refresh intervals covers this in detail.