How every number on this site is produced

Every vacancy is verified first-party at the employer, on the same page a candidate would apply through, and carries a dated audit trail: first seen, last confirmed, and the classification rule applied. Aggregator listings and recruiter duplicates are never admitted.

Each source is recorded with its employer, board type and board URL, and carries a permission flag; a source marked not permitted is never read. Every collection run records its status and the number of vacancies returned, so a silently failing source shows up as a gap rather than as a drop in demand.

Editorial firm profiles cover a wider set of employers than we read boards for. Where we have no live feed for a firm, its profile says so plainly rather than showing an estimated count.

Employer boards are re-read on a rolling schedule rather than all at once, so the dataset is continuously refreshed rather than rebuilt. Each source carries the timestamp of its own last successful read, and the most recent read across all sources is shown at the top of this page and beside the figures it supports.

A vacancy that has disappeared from its board at the next read is marked closed with that date. Closures are recorded, never deleted, which is what makes time-to-close and reopening measurable.

The compensation model is versioned separately and revised on a slower cycle: the current method is 2026.1, last revised September 2026. Every page that shows a modelled figure prints that version beside it, so a figure taken from the site can always be traced to the method that produced it.

A failed collection run is recorded as a failure. It never appears as a fall in demand, and figures that depend on a source which has not been read recently are labelled rather than quietly carried forward.

Employers routinely list the same role several times - once per office, once per level, or repeatedly as the requisition is refreshed. Counting those as separate demand would overstate every figure on the site.

Each vacancy gets a deduplication key built from the employer, its normalised title and its resolved city. Normalising a title strips seniority prefixes, requisition numbers, bracketed office names and punctuation. Two postings with the same key are one vacancy; the newest sighting updates the existing record rather than creating another.

A vacancy that disappears from its board is not deleted. It is marked closed with the date, which is what makes it possible to measure how long roles stay open.

Each vacancy description is matched against a fixed, published list of terms per skill. A mention means the employer asked for the skill in writing. We do not infer skills from job titles, from the firm's reputation, or from any model's judgement.

Terms and thresholds are set in advance and applied identically to every posting, so the same vacancy always produces the same result and any figure can be reproduced from the source link.

The consequence is a known undercount: an employer who wants C++ but never writes the word is invisible to this method. We prefer the undercount to a guess.

Two distinct kinds of pay figure appear on this site, and they are never mixed. Measured ranges come only from vacancies where the employer advertised a figure - currently 248 of 1,457 live vacancies. Those are parsed conservatively: annual figures only, with implausible values discarded, and reported as medians of the advertised minimum and maximum.

Figures are expressed as annual total compensation (base salary plus expected performance bonus) in the reference market's local currency, converted from a US dollar base. They describe typical market ranges rather than any individual offer, and exclude sign-on payments, deferred equity and carried interest. Portfolio-manager figures reflect formulaic profit-share arrangements and are therefore far more dispersed than any other role.

Modelled bands describe the market, not an individual offer. Where an employer states pay, the stated figure is used and labelled as such.

Every figure carries the number of vacancies behind it and the number of employers those came from. A confidence band is derived from both: high, medium, early signal, or insufficient. An early-signal figure is published but should not be read as a trend.

Period-over-period change is only published where the comparison period held at least five postings. Below that, the change reads 'insufficient' rather than showing a swing that is arithmetic noise.

We do not claim a trend from a handful of vacancies, and we do not round small samples up into confident language.

A vacancy enters the quant index only when its title or department matches one of seven published quant families: quantitative researcher, quantitative trader, quantitative developer, quantitative analyst, low-latency trading engineer, portfolio manager, and machine-learning researcher. Nothing is admitted on the strength of a generic word such as 'engineer' or 'developer'.

Quant firms hire far more widely than those seven families, and the broader roles they advertise - software, data, infrastructure, security, operations and corporate functions - remain fully searchable on the jobs index. Membership of the quant index is a specialism marker for measurement, not a statement that a role is less relevant.

Software and data engineering roles are tracked separately and counted in the quant index only where the advert itself carries trading-systems evidence, such as execution, market data, exchange connectivity, low latency or a trading platform. Without that evidence they are labelled engineering, non-quant, and are excluded from every quant figure on the site.

Each job page shows the classification applied to that vacancy: the family, the trigger found in the advert, whether it counts in the quant index, the seniority read and the resolved market. If a classification looks wrong, the rule and the evidence are both on the page, so it can be challenged.

This dataset measures advertised demand, not total hiring. A great deal of senior quant hiring - portfolio managers especially - never reaches a public board, so senior counts understate reality more than junior counts do.

Coverage is uneven by firm. Employers on boards we read are measured precisely; others appear only in editorial profiles. Comparing two firms' vacancy counts is only meaningful where both are read live.

Posting dates are the employer's own and are sometimes refreshed when a requisition is re-advertised, which can make an old role look new. Location fields are free text and resolve to a city only where the text is unambiguous; the rest are counted in totals but not in any city figure.

Advertised pay skews towards jurisdictions that require it, so measured pay ranges are not a representative sample of global compensation.