Metrics framework methodology · version TV-1.0

How the figures here are computed

Every formula, threshold and rule of this framework — generated from the same code that applies it, so this page cannot drift from the implementation.

Scope

This framework describes how retrieved data becomes metrics, what evidence supports them, and where they stop. It is applied to the individual markets this application already carries; it retrieves no additional sources for that.

It contains no buy, sell or hold recommendation and is not investment advice. A model value is an estimate under stated assumptions, an exposure index is an observation rather than a probability of loss, and confidence refers to data and model, not to a price development.

The data flags and flag points in the market lists follow the data flag methodology (currently version 2.7); this framework's exposure indices are built from the same flags. The evidence per dimension on the market pages (since 28.09.2026) is separate from this: it awards no points and feeds into no index.

Classification and method selection

The valuation method follows from the asset type. The mapping is given in full below and is made in exactly one place. An undetermined type gets no method — not the one that usually fits.

Source hierarchy

Every metric names where its value comes from. The tier follows from the kind of source and is not maintained per record, so no record can claim a tier its source does not support.

1. The issuer itself, a prospectus, a reserve report, the contract.
2. Chain, indexer, protocol interface, trading venue — places that hold the value themselves.
3. Data providers, aggregators, press.
4. Computed here from other values.
5. Origin not determinable — then the figure is no basis for a valuation.

When there is no figure

A missing metric is its own named state, never a zero. The difference between “the source does not carry this field”, “the series is too short” and “this model does not apply here” is the actual information.

DATA_NOT_AVAILABLE

Field not available

NOT_AVAILABLE

Not supplied

DATA_STALE

Value stale

SOURCE_UNAVAILABLE

Source unreachable

INSUFFICIENT_HISTORY

History too short

INSUFFICIENT_DATA

Inputs incomplete

MODEL_NOT_APPLICABLE

Model not applicable

NOT_APPLICABLE

Not defined for this asset

INSUFFICIENT_EVIDENCE

No evidence

CALCULATION_FAILED

Calculation undefined

UNKNOWN

Not checked

ORDERBOOK_DATA_UNAVAILABLE

No order book

VALUATION_NOT_SUPPORTED_BY_AVAILABLE_EVIDENCE

Valuation not supported by the available evidence

SCENARIO_PROBABILITIES_NOT_CALIBRATED

Scenario probabilities not calibrated

Formulas

Every metric with its definition, formula, inputs, reading, limit, method origin and status. No reading says whether a value is good or bad.

Exposure per dimension

Twelve separate dimensions, each an index from 0 to 100: how many of the flags checkable in it were triggered. The flags come from the same evaluation the market explorer uses — there is no second path from a market row to a risk figure.

Not a probability of loss, not a grade, no traffic light and no total. A dimension without a signal stays unchecked rather than becoming zero: zero would mean “checked, nothing found”.

How these indices relate to the dimensions of the market pages and the analysis framework →

Market regime

A descriptive placement of the time series into one of nine states. The rules are checked in this order because the conditions overlap; without a fixed order the result would depend on who checks them. No state is good or bad.

1. If there was barely any trading in the period (volume to liquidity below 0.01), that describes the state best — this rule comes before all others.
2. A confirmed departure from the look-back window's range, upward or downward.
3. Directional movement once the directionality measure reaches 25, where the direction comes from the structure of turning points rather than from the directionless measure itself.
4. Variation above 1.5 times or below 0.5 times the comparison set's median.
5. Otherwise a range — but only if both the directionality measure and the variation were available, so “range” does not become another word for “nothing measured”.
6. Otherwise: the inputs are insufficient for any rule. That state is not replaced by the most plausible one.

The thresholds are conventions, not empirically calibrated.

Findings on contract properties

Every finding follows a documented rule, and the severity follows a single consideration: how far does the capability reach, and is it bound to a delay or a multi-signature requirement? A capability is assessed, not an intention.

A property that could not be determined produces neither a finding nor a clean bill of health: it is reported as unchecked. A bound capability is placed one step lower but never becomes a mere observation — a bound capability is still a capability.

Dependency on the price source

The contributions add up to a dependency level of at most 100. If one of the three load-bearing inputs is missing — source count, fallback, what the price decides — no level arises: a figure from half the inputs gets read as checked later on.

• singleSource: +50
• fewSources: +25
• manySources: +10
• noFallback: +25
• criticalityLiquidation: +25
• criticalitySettlement: +15
• criticalityDisplayOnly: +0

Model confidence

The mean of the determinable inputs: field completeness, source tier, freshness, agreement between independent sources, length of history, and out-of-sample validation of the model. Inputs that cannot be determined do not enter as zero; they are reported as not determinable.

Independent of the model value and the exposure index: neither is an input to this calculation. High confidence alongside a low model value is therefore not a contradiction.

As long as no out-of-sample run is recorded for a model, that input is not determinable and depresses confidence. This is intended: no model in this framework counts as validated today.

Scenarios and simulation

Three scenarios, each with explicitly stated inputs. Probabilities are not invented: without empirical calibration none is shown, and no expected value can be formed. Three identical input sets produce no range, just the same result three times.

A simulation runs only with input distributions whose origin is documented, and it is reproducible: seed, number of runs, distributions and assumed relationships are reported with the result. If more than a tenth of the runs produce no result, no distribution arises — the surviving runs are not the model's distribution.

Validating the models

The validation infrastructure exists: growing and rolling time windows, a training set cut at the decision time, and a check that finds any point which crossed the boundary anyway. A run counts as evidence only if it contains no violations, has at least one evaluation window, carries enough prediction/outcome pairs, and was produced under the current method version.

Point forecasts are assessed by mean absolute error, root mean squared error, median absolute error and interval coverage; probability forecasts by precision, recall, area under the precision-recall curve, Brier score and calibration. The two families stay separate, and an index is not a probability. Every model is compared against the naive carry-forward.

How long a value holds

One window per kind of data, after which a retrieved value is reported as stale. Event-driven means: the value does not age with time but with a change — marking a contract architecture stale after two hours would spend attention where nothing changed.

MARKET_PRICE

5 minutes

MARKET_STRUCTURE

1 hour

ONCHAIN

6 hours

TOKENOMICS

168 hours

FUNDAMENTAL

720 hours

LEGAL

event-driven

CONTRACT_ARCHITECTURE

event-driven

What this framework cannot do today

• No price series per token. What is available is a yield and capital series per market; every technical metric refers to that and not to a price.
• No order book. Spread, depth, price impact and slippage are therefore not reported and not approximated.
• No holder distribution. Concentration measures and the attribution of pooled addresses stay undetermined.
• No supply register: total supply, emissions, burns and unlock schedules are missing.
• No contract data. Every property this framework checks counts as unchecked, and no security finding follows from that.
• No net asset values, bond terms, reserve reports or protocol revenue. This is why no model value arises for any asset in this application today.
• No model in this framework has been validated out of sample. The infrastructure exists, the runs do not.
• No legal classification. None is inferred and none is claimed.

Versions and changes

Every change to a model, formula, weight, threshold, data source or classification is recorded with its date, the person responsible, the old and new value, the reason and the version.

• 2026-09-25 · MODEL · TV-1.0

  token-valuation-framework

  no valuation framework → TV-1.0

  Introduction of the token valuation framework: calculation engine, evidence binding per metric and a vocabulary of failure reasons. No asset is valued, because the input data for that is not available.

  Responsible: Tobias Licht

• 2026-09-30 · DATA_SOURCE · TV-1.0

  dimension-limits-evidence

  Limits per dimension without the evidence layer → Limits per dimension name the evidence on the market pages

  Since 28.09.2026 the market pages show evidence from public sources (DeFiLlama, GLEIF, SEC EDGAR) for smart contracts, oracles, governance, counterparties and operations. The description of the limits still said such data was unavailable. The indices and their calculation are unchanged; evidence feeds into no index.

  Responsible: Tobias Licht

The register begins on the day it was introduced. Before that, nothing stands here on purpose: a retrospectively assembled list would be a claim about a past nobody checked.

To the analysis framework →To the data-flag methodology →To the market explorer →