Skip to main content
Skip to content

CryptoIQ

Methodology

How every CryptoIQ score is built. The same framework runs across our scores — the specific signals, thresholds, and weights stay proprietary. Exit IQ works differently and is covered at the end.

What our scores are

Every CryptoIQ product — Cycle IQ, Altcoin IQ, Attention IQ — is a single 0–100 read on where a market sits in its cycle. Each is a regime classifier — not a price target, not a trading strategy, not a buy or sell signal. It describes what phase a market appears to be in, not what to do about it.

Each score maps 0–100 onto five labeled bands, running from one regime extreme to the other. The band names differ by product; the 0–100 shape is shared.

  • 0–20
  • 20–40
  • 40–60
  • 60–80
  • 80–100

How we build the models

Labeled history, binary signals, one composite.

Labeled ground truth. We don't eyeball tops and bottoms. The cycle events each model is measured against are labeled by a fixed rule — a triple-barrier method (López de Prado, Advances in Financial Machine Learning, 2018) — not by hindsight.

Binary signals across categories. Each input is a binary fire / don't-fire signal with a fixed threshold, drawn from categories such as on-chain, market structure, sentiment, macro, and — for token products — fundamentals. Correlated signals within a category are combined so nothing is double-counted.

Weight-of-Evidence composite. The signals are blended through a Weight-of-Evidence composite into a calibrated 0–100 score. The specific signals, thresholds, counts, and per-category weights are proprietary.

How we validate

How we know a score holds up.

Event-labeled ground truth. Performance is measured against the labeled event set above, not against a story told after the fact.

Per-band Wilson confidence intervals. Every accuracy figure is reported with its 95% Wilson interval (correct for small samples). A point estimate without an interval is misleading, so we don't publish one.

Permutation tests. To rule out base-rate-only performance, each figure is compared against the distribution from reshuffling labels at random thousands of times. If a model can't beat random labels, the claim doesn't ship.

Proprietary and versioned

The framework is public; the recipe is not. The specific signals, thresholds, and weights stay proprietary — that is the product.

When the methodology changes, it ships as an explicit, versioned update with fresh validation — never a silent, continuous refit. Earlier versions stay on record, so any past score can be read against the model that produced it.

Limitations

What these scores cannot tell you.

A regime read, not a buy signal. A score answers “is this market euphoric or fearful?” — not “should I buy or sell?”. Position sizing, entry timing, and risk tolerance are not in any model and never will be.

Bottoms are zones, not dates. Markets grind sideways for months and rarely produce a single clean capitulation print. Treat the low-end bands as zones, not precise calls.

Markets are non-stationary. Cycle compression, regulatory shifts, and changes in market structure can degrade a signal that worked historically. We do not assume the future repeats the past.

Past performance is not predictive. Every statistic describes historical fit, not a forecast. Read the numbers as descriptive, and see the disclaimer.

Exit IQ

Direct wallet tracing, not a composite score.

Exit IQ is built differently from the scores above. There is no Weight-of-Evidence composite. We trace a token's team-treasury and market-maker wallets on-chain and report when they move supply to exchanges, where it can be sold. It reports the transfer; it does not infer intent.

Two coverage tiers. Traced tokens have their team, treasury, and market-maker wallets individually attributed, so the read is who moved supply and where it went. Screenedtokens show total exchange flow only — we can see supply moving, but not yet who. A token stays Screened until we have traced its wallets.

The balance checksum. For every traced wallet we replay its full transfer history, and the result must equal the wallet's live on-chain balance. That match is how we confirm a scan is complete; a wallet that fails it is not published as traced.

What it cannot tell you. Coverage is Ethereum mainnet only, so deposits made on other chains are not yet captured and multi-chain tokens are undercounted. Wallet attribution is our own on-chain inference — it can be wrong, and when it is, we correct it publicly. And the read can lag price turns: it describes where supply is flowing, not when the market turns. Not investment advice; see the disclaimer.