Salary Scout

Posting

Load an example

Each block can be switched off to see how the estimate changes without it. The model was trained to cope with missing blocks, so a masked block is not "unknown", it is genuinely absent.

Titleuse Job title Core title the role stripped of level and specialty, as the source site labels it

Descriptionuse Description paste the posting; HTML and salary figures are stripped before hashing Requirements summary optional, a sentence or two

Role metadatause

Seniority Job category Years of experience Management years Commitment Security clearance Bachelor's Master's Doctorate Visa sponsorship

Technical tools comma separated

Locationuse

Workplace type Primary state Number of states listed Countries comma separated codes Latitude Longitude

Companyuse Employer employers seen in training get their own pricing history; others get the average

Employees Year founded Organisation type Sector HQ country Applicant tracking system

Industries comma separated

Loading the model (about 7 MB)…

Estimated advertised range

—

Midpoint —

Typical error for these inputs: —
Ranges that overlapped the real one: —

What moved the estimate

Exact Shapley values over the five blocks, relative to the average posting in training. Blue raises the midpoint, red lowers it.

About this model

A gradient-boosted model (LightGBM) trained on about 20,000 US postings that disclosed a salary, published April to September 2026. Text is hashed into bag-of-words features, so no vocabulary is shipped; employer identity is a target-encoded lookup. The five blocks were randomly dropped during training so the model behaves sensibly when you switch them off.

Errors quoted above are measured on the newest 15% of postings, which the model never saw. The estimate is what a disclosing employer would advertise, not what a specific person would be offered. Individual-contributor roles only, mostly data, analytics and software.

Source, report and methodology.