Methodology

How Noble Xenon works

Which markets exist, how they resolve against the Noble FX Fix, how forecasters are scored, and how the consensus is built. Everything here is applied the same way to every forecaster, and every published figure records the methodology version that produced it.

Current methodology versions
Component	Current version
Noble FX Fix	fix-1.0.0
Scoring and credibility	scoring-1.0.0
Consensus	consensus-1.0.0

The consensus is a credibility-weighted view of what forecasters submitted. It is not a forecast by Noble, and nothing on this page or the Service is financial advice.

Markets

A market asks one question: where will a currency pair’s Noble FX Fix be on a given date? Forecasters answer with a single rate.

Pairs. 14 currencies, each quoted against the US dollar: AUD, CAD, CHF, CNH, DKK, EUR, GBP, INR, JPY, MXN, NOK, NZD, SEK, SGD. Each pair is shown in its market-standard quotation (for example EUR/USD and USD/JPY). Pegged currencies are not listed.

Open for forecasting. Forecasting is open on 6 pairs: AUD/USD, EUR/USD, GBP/USD, USD/CAD, USD/JPY, USD/MXN. The Noble FX Fix is published for all 14 pairs.

AUD/USD
USD/CAD
USD/CHF
USD/CNH
USD/DKK
EUR/USD
GBP/USD
USD/INR
USD/MXN
USD/NOK
NZD/USD
USD/JPY
USD/SEK
USD/SGD

Resolution dates. Markets are created for the end of this week, the end of next week, the end of the month, the end of the quarter, the end of the year, and a rolling ladder of quarter-ends. Each resolves on the last trading day of its period: weekends and US market holidays are skipped.

One market per date. When several periods end on the same day, they share one market. For example, the last trading day of December can be the end of a week, a month, a quarter and the year at once. A rate on a date is one fact, so it has one market.

Lifecycle.

Open: Forecasts are accepted. You can forecast the same market more than once.
Closed: Forecasts stop at 15:00 London on the resolution date, 60 minutes before the 16:00 London fix.
Resolved: Against the Noble FX Fix (section 02).
Voided: If no fix is available within the roll window (section 03).

Forecast checks. A forecast must be a positive rate with at most 6 decimal places. When a current spot rate is available, it must also be within 50% of it. A forecast outside that range is rejected with a suggested correction, because it is almost always a typing slip or an inverted quote.

The Noble FX Fix

Every market resolves against the Noble FX Fix: the mean of 6 spot prints taken at 3-minute intervals over the 15 minutes to 16:00 London, on trading days only.

Fix print times, London
Print	London time
1	15:45
2	15:48
3	15:51
4	15:54
5	15:57
6	16:00

Because the prints are evenly spaced, their simple average is a time-weighted average price (TWAP) over the window. Times are London wall-clock times, so the fix follows UK daylight saving automatically.

\text{fix} = \frac{1}{n}\sum_{j=1}^{n} p_j, \qquad n \ge 4

Minimum prints. A full fix needs at least 4 of the 6 prints. If only 1 to 3 prints were captured, the fix is the last print at or before 16:00 London, and it is labelled degraded wherever it appears. A degraded fix is never described as a TWAP. If no prints were captured for a pair, no fix is published for that pair that day.

Transparency. Every raw print is kept permanently, and each fix records how many prints it used; the count is shown with the resolution on the market page. Spot data comes from a third-party market-data provider.

Not a licensed benchmark. The Noble FX Fix is aligned with the WM/R 4pm London fix methodology, but it is not the licensed WM/R rate or any other regulated or licensed benchmark.

Resolution

A market resolves against the first Noble FX Fix published on or after its resolution date: normally the fix on the resolution date itself, shortly after 16:00 London.

Roll-forward. If no fix was published on the resolution date (for example, because no prints were captured), the market resolves against the next published fix, up to 3 trading days later. The market page shows the fixing date actually used and flags it when it differs from the resolution date.

Voided markets. If no fix is published within those 3trading days, the market is voided. Its forecasts are not scored and do not count toward anyone’s record. We void rather than resolve against a rate that may be materially out of date.

Resolutions are final. Each resolution records the rate, the fixing date it used and the fix methodology version. Once a market has resolved, it is never re-resolved and its resolved rate never changes.

Scoring and credibility

Each forecaster earns a credibility score between 0 and 1, built from how accurate they have been. Accuracy dominates; consistency can adjust it slightly but can never rescue an inaccurate record.

Relative error. Every resolved forecast is measured as a percentage of the resolved rate, so a miss on USD/JPY and a miss on EUR/USD are directly comparable:

\text{error} = \frac{\lvert \text{forecast} - \text{fix} \rvert}{\text{fix}}

Horizon-scaled accuracy. Forecasting a week out is easier than forecasting a year out, so the tolerance for error widens with the time between submission and resolution, in line with how exchange-rate volatility grows with the square root of time:

\text{accuracy} = \exp\!\left(-\frac{\text{error}}{0.005 \cdot \sqrt{\max(\text{days},\, 1)}}\right)

where days is the number of calendar days from submission to the resolution date. Each forecast is scored against its own horizon, and your accuracy is the average of those scores across all your resolved forecasts. Every forecast you make on a market counts, not just your last one, so being right early earns more than converging late.

Short and long horizons. Forecasts made 30 days or less before resolution are scored in the short bucket; longer ones in the long bucket. Accuracy, consistency and credibility are calculated separately for each bucket, and the two are ranked separately.

Consistency. If you forecast the same market several times, the spread between your highest and lowest forecast, as a share of the resolved rate, is a reliability signal. A market you forecast once has a spread of zero. Your consistency in a bucket comes from the average spread across the markets you forecast in it:

\text{spread} = \frac{\max - \min}{\text{fix}}

\text{consistency} = \exp\!\left(-\frac{\overline{\text{spread}}}{0.05}\right)

It enters only as a bounded multiplier, so it can lower credibility by at most 20%:

\text{credibility} = \text{accuracy}_{\text{shrunk}} \times \left(0.8 + 0.2 \times \text{consistency}\right)

Earning influence over time. Measured accuracy is blended toward a neutral 0.5 until a record builds up, so a single lucky forecast cannot buy maximum credibility:

\text{accuracy}_{\text{shrunk}} = \frac{n \cdot \text{accuracy}_{\text{measured}} + 4 \cdot 0.5}{n + 4}

where n is the number of resolved forecasts in the bucket. A forecaster with no resolved forecasts in a bucket starts there with a credibility of 0.5.

Leaderboard. Forecasters appear on the leaderboard once they have at least 5 resolved forecasts in a bucket. Statistics are rebuilt in full from every resolved forecast after each daily fix, never updated incrementally, and rank history is recorded at the same time. Because they are rebuilt in full, scores always reflect the current scoring version; the resolved rates they are measured against never change.

Consensus

The consensus for a market is a credibility-weighted average of forecasters’ latest views. How much of it is published depends on how many forecasters’ forecasts pass the sanity gate below:

Full consensus from 8 forecasters.
Early read from 3 to 7forecasters. The value is shown, but marked “Early read” with the forecaster count and drawn more lightly, because a handful of forecasters is a noisy signal.
Collecting below 3forecasters. The market shows as “collecting” and no value is shown.

The calculation is the same in every tier; the thresholds only decide what is shown.

Eligible forecasts.Only each forecaster’s most recent forecast on the market is used, because the consensus answers what the crowd thinks now. (Scoring, by contrast, uses every forecast.)

Sanity gate. Forecasts more than 25% away from the current spot rate are excluded. The number of excluded forecasts is recorded with each consensus and shown alongside it on the market page, never hidden.

Weighting.Each eligible forecast gets two weights multiplied together: the forecaster’s credibility in the horizon bucket of that forecast (who you are), and a proximity weight that fades forecasts far from the median of all eligible forecasts (how far from the crowd this forecast is):

w_i = \text{credibility}_i \times \exp\!\left(-3 \cdot \frac{\lvert f_i - \text{median} \rvert}{\text{median}}\right)

The consensus is the weighted mean of the forecasts using those weights:

\text{consensus} = \frac{\sum_i w_i f_i}{\sum_i w_i}

Band. The band around the consensus is one weighted standard deviation either side. It measures disagreement among forecasters, not uncertainty about the outcome. Forecasters can agree closely and still be collectively wrong.

\text{band} = \text{consensus} \pm \sqrt{\frac{\sum_i w_i \left(f_i - \text{consensus}\right)^2}{\sum_i w_i}}

Confidence.The average credibility of contributors, weighted by each contributor’s actual influence on the consensus:

\text{confidence} = \frac{\sum_i w_i \cdot \text{credibility}_i}{\sum_i w_i}

When it updates.A market’s consensus is recalculated each time a forecast is submitted to it, and every market that has not yet resolved is recalculated after each daily fix. If no current spot rate is available, that recalculation is skipped until the next one. Every calculation is stored with the spot rate used and its methodology version, so past figures can be reproduced. Individual forecasts are shown anonymously on the market page; credibility is shown by dot size.

Versions, changes and corrections

Versioning. The fix, scoring and consensus calculations each carry a version number (current versions are listed at the top of this page). When a calculation changes, its version changes. Every fix and every consensus calculation is stamped with the version that produced it and keeps it. Forecaster statistics are rebuilt in full after each fix, so they always carry the current scoring version. Material changes are announced on this page with the date they take effect.

Corrections. Published fixes are final and are not restated, because forecasts have already been scored against them. If a published fix is found to contain a material error, such as bad source data, we publish a dated notice here naming the fixing date and affected pairs. Scores stand. Degraded fixes, rolled-forward resolutions and voided markets are the methodology working as designed, not errors.

Change notices

2026-10-22Early-read tier. A market’s consensus is now also published once 3 forecasters contribute, as an early read: shown distinctly and labelled “Early read · n=N”, with a lighter band. The full consensus is still published from 8 forecasters. This changes what is published, not how the consensus is calculated, so the consensus version is unchanged.

Correction notices

No corrections have been issued.

Limitations

The consensus reflects the forecasters who chose to take part. It is not a sample of the market and not a prediction by Noble.
At low participation, a few forecasters can move the consensus significantly. The 3-forecaster floor and the “early read” label below 8 limit this but do not remove it.
Spot data comes from a third-party provider and can be delayed or wrong. The fix inherits any error in the prints it is built from.
Scoring parameters, such as the volatility scale, are global today. Some pairs move far more than others, and per-pair calibration may be introduced in a future version.