Methodology

01 The record is the method

Every call is timestamped before the outcome. Major entry facts remain fixed. Prediction revisions retain dated history. Each graded result resolves on the call's permanent URL.

When the measurement itself improves, the record says so. Grading methodology changes are marked as dated era boundaries and the eras are reported separately, rather than restating history under new rules.

published

before the outcome, timestamped, immutable

graded

net of modeled execution costs, not headline moves

eras

methodology changes marked, never backfilled

02 Evidence before opinion

Every cycle, each covered market gets a structured dossier assembled from source classes: derivatives positioning and funding, DeFi flows, macro regime, on-chain network stress, exchange order-flow tape, options surface and skew, historical regime analogues, and a probabilistic price forecast.

That list is what the desk reads today, not what it has ever read. A labeled exchange-flow leg ran on a paid vendor through the calibration window and was retired in August 2026 when it had still never fed a trading decision. It is named here because a source list that quietly drops entries is not a record.

The dossier is compiled before any model reasons about direction. Opinion formed without evidence is not permitted into the pipeline; a source that fails to load is reported as missing, not silently imputed.

Coverage is governed, not opportunistic. A market enters the traded universe only once a venue we already ingest can serve its minute-level OHLCV as a primary source, because every reading above is computed from those bars. Most markets clear that gate through Coinbase; the ones it does not list name their own venue on the dossier, so a differing data source is disclosed rather than hidden.

03 Probabilistic forecasting

The desk ran a time-series foundation model until July 2026. It was graded against naive baselines, it did not beat them on direction, and it was retired rather than kept for appearances.

What replaced it forecasts uncertainty instead of direction. An exponentially-weighted volatility estimate gives every horizon a p10 to p90 band around the current price, and the point forecast sits flat inside it. The band sizes the risk. It does not claim to know the way, and no part of the pipeline reads a direction out of it.

04 An adversarial ensemble

Theses are argued, not asserted. A panel of role-specialised analysts, drawn from distinct model lineages, reaches independent verdicts over the same dossier. Votes are weighted by each analyst's measured track record, and a well-earned dissent can override the room.

Consensus, dissent and conviction are all recorded. When a settled outcome proves the dissent right, that analyst's standing rises, and its next vote counts for more. Reputation is earned from graded results, never assigned.

05 Discipline between verdict and position

A verdict is not a position. The projected edge has to clear estimated costs and uncertainty before capital is modeled into the trade; setups that fall short are logged as deliberate passes. Most reviewed setups end there, on the record.

Sizing is fractional Kelly, clamped by hard per-position and correlation-cluster caps, under a peak-to-trough drawdown pause that halts new risk without disabling the desk. Positions open with volatility-derived stops and targets and are monitored tick by tick, wick-aware, until close.

Every committee verdict is graded, whether or not capital follows it. The desk keeps two ledgers: the committee ledger records each directional call at its published levels, and the book ledger records only the calls that were actually sized into the paper book. When exposure caps leave less room than a position is worth opening, the size is trimmed to fit; when the remainder is too small to be worth the costs, the call is logged as not sized and graded against what it would have done, net of estimated costs. A call the book declines still counts on the committee's record. Capital discipline never edits the scorecard.

entry

edge must clear costs and uncertainty, or pass

sizing

fractional Kelly, capped, correlation-aware

risk

drawdown pause, volatility brackets, continuous monitoring

dual ledger

committee calls graded; book fills sized within caps

06 Grading and calibration

Trade calls are graded at their published levels, net of modeled fees, funding and slippage, so the record reflects what execution would have cost, not a frictionless fantasy.

Prediction calls are scored on Brier against the market's own price at call time. Alpha means beating the market's Brier score, not a raw hit rate: being right 60% of the time when you said 60% is calibration; being right more often than the price implied is edge. Every published rate carries a Wilson 95% interval and expectancy carries a bootstrap interval, so small samples read as small samples.

majors

net-of-costs P&L at published levels

predictions

Brier vs the market's implied probability

uncertainty

Wilson intervals on rates, bootstrap on expectancy

07 Where the numbers come from

Every reading on this site is computed from minute-level exchange candles the desk ingests itself and stores in its own time-series warehouse. Spot OHLCV and the trade tape come from a primary exchange venue per market, with named alternates behind it. Derivatives funding and open interest come from perpetual venues directly. Nothing on the trading path depends on a price aggregator.

The remaining source classes are public endpoints: options surface, DeFi totals, macro series, and chain-level network stress. Where a market's spot data comes from a venue other than the default, its dossier says so on the card rather than presenting one venue's bars as another's.

The desk buys reliability, not latency. A faster feed would not change a decision cadence measured in minutes, so paid tiers are bought for rate-limit headroom and coverage when a free tier stops being sufficient, and for nothing else. What each class costs is an operational detail; what it feeds is on this page.

spot + tape

exchange venues, ingested and stored locally

derivatives

perpetual venues, read directly

disclosure

a non-default venue is named on the card

08 What we do not claim

The desk currently trades on paper with modeled execution costs, disclosed as such. We do not promise future returns, and we do not present short samples as proof: where the n is small, the intervals are wide and we say so.

The method itself is under permanent measurement. Calibration experiments run continuously, and changes to how the desk decides ship only after the evidence reads, with the boundary marked on the record. Everything published here is research, not financial advice.