Kinzan 金山 Journal
Menu

Methodology

Revised 2026-10-08

How the figures on Kinzan are calculated, and the limits each one carries.

The 24-hour move

change24h is the percentage change over the trailing 24 hours, measured within one venue's own series. Direction uses a ±0.05% deadband: inside it a move is shown as flat. The 24-hour high and low are taken from hourly closes, so they are not the true intraday extremes. Range position is where the price sits between that low (0%) and high (100%).

Every move carries a ▲ or ▼ glyph and a signed number as well as a colour, so direction never depends on colour alone.

Volatility (sigma20)

sigma20 is the sample standard deviation of an instrument's last 20 daily returns, in percent — the size of move it makes on an ordinary day. Returns are measured across non-trading days: Friday to Monday is one return, not three. A missing day is left missing; it is never filled with the previous close, because doing so to every weekend would understate a currency's volatility by roughly 40%. With fewer than 15 returns, or a volatility of exactly zero (a pegged currency), no figure is published.

"Vol vs baseline" compares today's hourly volatility with the instrument's own exponentially weighted average, which has a memory of about three days. It is only shown once the baseline has at least twelve samples.

Surprise

surprise is the 24-hour move divided by the instrument's own 20-day volatility, signed, in standard deviations (σ). It is the one figure on the site that compares instruments with each other: 0.4% in USD/JPY against a 0.16%-a-day habit is +2.5σ, while 2% in Bitcoin against a 2%-a-day habit is +1.0σ. Those two readings mean the same thing, which no pair of percentages does.

One limit travels with every surprise reading: its numerator is a trailing 24 hours while its denominator is close-to-close over whole days, so mid-session the reading runs slightly tame.

Phrases such as "the kind of session it has roughly once a month" convert a reading into a frequency using a normal distribution. Real market returns have fatter tails than a normal distribution, so large days happen more often than these phrases imply.

Momentum lenses

Each momentum page ranks a fixed population — currencies, metals, crypto, commodities, the headline "majors", the whole board, or the whole board by surprise. Membership is defined by asset class rather than a hand-kept list. Ranked across the whole board by percentage, a currency pair would almost never appear, because exotic pairs and coins out-move the majors on nearly every day; ranking within a class keeps the comparison like-for-like.

Currency strength

For each currency pair the base currency takes the pair's 24-hour move and the quote currency its negative. A currency's strength is the unweighted mean of its contributions. It is not a trade-weighted index: coverage is uneven, so the number of pairs is shown with every reading.

The currency race stretches the same calculation over the daily window. Each pair is measured as its percentage change from its own close at the start of the window, and votes that change on its base currency and the negative on its quote. A day is included once at least half of the pairs traded it; on a day a pair did not trade, its last close stands, because the race tracks a level rather than a daily return. A pair with no close within three days of the window's start sits the race out, and a currency with no pair voting on a day has no bar that day rather than a bar at zero.

Sessions

The day is divided into three contiguous UTC blocks — Asia 22:00–07:00, London 07:00–13:00, New York 13:00–22:00 — so each hour is counted exactly once. Each hourly step is attributed to the block its later candle falls in, and the steps are summed as log returns, so the three session figures compose exactly to the day's move. These are price moves only; they say nothing about volume.

Correlations

Correlations are Pearson correlations of daily returns over the last 60 sessions that both instruments traded — never of price levels, and never across a day only one of them traded. At least 30 paired returns are required. "Relationships that have changed" compares the 60 sessions before the last 20 with the last 20; the windows are adjacent, not nested, so a recent break cannot contaminate its own baseline. Correlation is not causation and not a hedge ratio.

On an instrument's page each relationship is shown on three windows (the last 60, the last 20 and the 60 before those 20) with a read in words, from fixed bands on the strength of the correlation (its size, whatever its sign). Holding: the last 20 within 0.2 of the 60 before. Strengthening or weakening: up or down by 0.2 or more. Flipped: at least 0.4 one way before and at least 0.2 the other way now. No real link: under 0.2 on both. The last 20 needs at least 12 paired returns.

The rolling correlation on a comparison page is one point per shared session: the correlation of the 20 paired returns that ended on that day. The first 19 shared sessions have no point rather than a thinner one.

Against its own range

A level is placed against the instrument's own earlier values in the daily record, about four months. Its percentile is the share of those earlier values below it, an equal value counting half, so a new high for the period is the 100th and a level that has not moved is the 50th; at least 40 earlier values are required. The strip runs from the lowest to the highest value in the window, with bands for the middle 80% and the middle half and a tick at the median. On an instrument's page the same is done for its 20-day volatility (the sample standard deviation of the 20 daily returns to that day, as the board computes it) and for its distance from the mean of its own last 20 closes. On the macro backdrop the curve is the US 10-year yield less the 13-week bill yield, on the days both closed. A high percentile says where a market sits for the season, not for the year, and it is not a fair value or a forecast.

Heat map

Each cell is the unweighted average return of the instruments in a class or group, over 24 hours (live) or 5, 21 and 63 trading sessions (from daily closes). It is not an index. The number of instruments behind each cell is shown, and tints are relative to the largest reading on the same grid.

Risk dial and market stress

The risk dial combines equity indices, major crypto, copper, the VIX, gold, the US Dollar Index and the strength of the yen, franc, Australian and New Zealand dollars. Weights are inverse to each input's typical daily range. The sum is passed through a hyperbolic tangent to give 0–100; 65 and above reads risk-on, 35 and below risk-off. Market stress is the share of instruments with a volatility figure whose move is beyond ±2σ; about 5% would be expected on an ordinary day.

Breadth

Breadth asks whether a move is broad or carried by a few names. An instrument is an advancer if its close-to-close return was up by more than the ±0.05% deadband, a decliner if down by more, and unchanged in between. The advance/decline line is advancers minus decliners, summed day by day from the start of the window. "Above its 20-day average" is the share of instruments whose close is above the mean of their own last 20 closes, and a 20-day high or low is a close at the highest or lowest of those 20.

A day is included once at least half of the instruments traded it, so a weekend, when only crypto trades, does not swing every series. An instrument that did not trade on a day is left out of that day rather than counted as unchanged, and one with fewer than 20 traded closes is not measured. An asset class is shown separately once at least four of its instruments are measured. Where a page states how many instruments are up or down, the count is over exactly the rows that page shows.

Market heartbeat

The heartbeat is breadth over the last 24 hours, hour by hour: advancers minus decliners as a share of the instruments measured, from +100 (every one higher) to −100 (every one lower). At each hour an instrument is compared with its price at the start of the window, using the same ±0.05% deadband. An hour in which fewer than 60% of instruments can be measured is left as a gap rather than drawn from a handful of series. The line is built from hourly closes, so its last point is close to, but not exactly, the board's own figure. The pulse rate follows the risk dial's distance from neutral in either direction; it is a visual mapping, not a measurement.

Volatility terrain and market quakes

The terrain draws each instrument's daily move over the window in its own standard deviations: the close-to-close return that ended on that day, divided by the sample standard deviation of the 20 returns before it, with at least 15 required. Days of 2.5σ or more are drawn in colour. A day the instrument did not trade is a gap in its line, never a flat day, and a day is drawn once at least half of the instruments can be measured.

A quake is a move of 3σ or more. Quake records are kept only for days that have finished in every market, and each record is written once and never revised, so a quake's page always shows the figure it was recorded with. The homepage switches to the terrain during a live quake: a 4σ move in a major instrument, any 5σ move, or a cluster of them at once — and only while that instrument's own latest price is recent.

Market constellation

Each instrument is a star. A link between two is their correlation, calculated exactly as on the correlations page: daily returns over the last 60 sessions both traded. A link is drawn at |r| of 0.5 or more, at most three per instrument with the strongest first, and any pair at 0.8 or more is always drawn. An instrument with no link that strong is tethered faintly to its strongest partner at 0.25 or more, so a weak relationship reads as weak rather than as none. Clusters are groups of three or more connected by positive links; moving in opposite directions is a relationship, not membership of a group. A star's size and depth are its |surprise|; an instrument with no surprise figure sits at the back, drawn dashed.

The replay steps through up to 40 past days. Each frame draws a pair's correlation over its last 20 sessions to that day, against the 60 sessions before them, and a break is the same test the correlations page uses: a usual |r| of at least 0.6 that has changed by at least 0.4.

World market clock

Each stock index stands on its exchange's city as a spike whose height is its |surprise| in σ. Each currency is a ring on its central bank's city, sized by its currency strength — a percentage, not a σ, so the two are never compared. Crypto has no location and orbits instead, sized by |surprise|. An instrument with no figure is drawn as a short dashed stub, not a zero. Exchange hours are the scheduled regular session on weekdays; holidays and half days are not known, so the clock says "scheduled open", never "trading now". The replay is built from hourly closes, each instrument rebased to the start of the window, so its last frame is close to the live figure but not identical to it.

Positioning and the crowd X-ray

Positioning comes from the weekly Commitments of Traders report: the speculative crowd's net position in each futures market, long contracts minus short. Its percentile is the share of that market's previous weeks at or below this week's figure, and needs at least 26 earlier weeks; a percentile against fewer would mean very little, so none is shown. 90 or above reads "near its longest", 10 or below "near its shortest". The crowd X-ray's time-lapse computes every earlier week the same way, so it ends on the figure the table prints.

Three limits travel with every reading. It is futures positioning, not the spot market, and futures are a small part of currency trading. It is a Tuesday snapshot published on Friday, so it is already several days old. And it is a level, not a signal: an extreme says the crowd is heavily one way, not what happens next.

Market map and bubble screener

No single measure of size covers every market, so each is sized within its own asset class: crypto, stocks and funds by market capitalisation, currency pairs by their share of trading turnover, and metals, commodities, indices and rates by a fixed relative weight, because a futures contract's volume belongs to the contract and an index has no size of its own. Each class then takes a share of the picture in proportion to how many instruments it holds. A tile's area therefore says how large an instrument is within its class; it does not say that gold is larger than a given company. The other sizing views — the size of the move, surprise and volatility — compare instruments directly, because each is measured within the instrument's own series. An instrument without the figure is left out of that view rather than drawn at zero. Trading volume is not offered, because it is not available on the same basis for currencies.

Trading hours

"When does it move" is the average absolute hourly return in each UTC hour. Absolute, because direction cancels: a market that rises 1% at 08:00 on half its days and falls 1% on the rest would otherwise read as the calmest hour of the day. The profile builds up over time, and an hour with fewer than five observations is shown as no data rather than as a quiet hour.

Chart indicators

Indicators are calculated from closes only — on daily charts the days an instrument traded, never a filled-in day: simple moving averages of 20 and 50 closes, a 20-close exponential average, Bollinger bands at two standard deviations either side of a 20-close average, and a 14-period RSI with Wilder's smoothing. Fibonacci levels span the high and low of the period on screen. Hourly charts use shorter windows (6 and 12 hourly closes; RSI 14). An indicator is not drawn until it has its full window. Indicators that need each day's open, high and low, such as the average true range, are not offered.

Economic calendar

Releases show their forecast and previous figure. The actual result is filled in only where it can be republished — currently US federal statistics — and stated as the release announces it, usually a change rather than a level. Where there is none, a past release reads "not yet reported" rather than carrying an estimate.

How the Alpha Brief is scored

Every long and short in the Alpha Brief is a call with three parts fixed when it is written: the price it was written against, a horizon (the next full trading day, or five), and a "wrong if" level on the losing side of that price. The level must sit at least one and at most four times the instrument's 20-day volatility from that price, scaled by the square root of the horizon in sessions: never inside one ordinary move for the horizon, where routine noise would end the call, and never so far away that the call cannot be proven wrong. A call without a level that passes is discarded by code before the brief is published, not repaired. "Watch" ideas are not calls and are not scored.

Calls are resolved on daily closes in the instrument's own series. The first session counted is the first close dated after the day the brief was written, never that day's own close, because some markets have already closed that day by the time a brief is written. A close at or beyond the "wrong if" level within the horizon ends the call there: it is stopped out, scored as a miss, and its result is that close, which can be beyond the level when a market gaps or falls hard within a session. Otherwise the call is marked on the horizon's last close: right if the price moved the call's way, wrong if against. Because only closes are kept, a level crossed during a session and recovered by its close does not end a call. A close is counted only once its day has finished everywhere. A scored call is never rewritten.

The hit rate is shown only from 20 scored calls, and always beside the rate of the same instruments over the same horizons on every day in the window: how often a call with no view at all would have been right. The live figures on a call (the move since, the distance to its level) use the price now and are not part of the record.

Base rate in the Alpha Brief

Each setup in the Alpha Brief carries a base rate computed by fixed arithmetic when the brief is written, not by the model. The instrument's move at that moment, in σ (its surprise), is put in a band: under 0.5σ either way, or 0.5–1σ, 1–2σ, or 2σ and over in the same direction. Every earlier day in the daily record whose own σ move falls in the same band is matched, and the next session's close-to-close return is counted. For a long the count is the next sessions that closed higher; for a short, lower. The same count over every day in the window is shown beside it, because a rate means little if the instrument rose on most days anyway. Below 6 matching days no rate is given. The next session is only counted once it has finished. A base rate describes the past few months; it is not a forecast and not a recommendation.

What Kinzan does not do

Kinzan describes what markets have done. It does not forecast, recommend or advise, and nothing on it is an offer to buy or sell anything. See the disclaimer.