Data Job Market Intelligence

Key findings

From 2,367 postings across 5 markets and 6 sources, accumulated by a weekly scheduled collector. Directional, not definitive — they sharpen as the data grows.

Python + SQL

lead demand in every market (19% and 14% of all postings) — the non-negotiable core for data roles.

$112k → $162k

median advertised salary rises Analyst → Scientist → ML Engineer, normalized to USD/year.

5 markets

Remote-international, US, Australia, Malaysia & Indonesia — each tagged and comparable.

🌏 Indonesia vs. remote-international

The markets want different things — useful if you're deciding which to target.

Indonesia (local, n=33)	Remote-international (n=1,362)
SQL 61%, Python 58%	Python 28%, SQL 19%
Tableau 30%, Power BI 27% BI-heavy	AWS 18%, Git 13% cloud/eng
Statistics 33%, ETL 27% analyst core	Machine Learning 14%, Kubernetes 11%

Indonesian roles lean analyst / BI (SQL + a BI tool are near-mandatory); remote roles lean Python + ML + cloud.

Charts

Generated reproducibly from the warehouse (python -m src.figures).

Most in-demand skills across all postings.

Median salary by skill — Median advertised salary by skill (advertised only, normalized to USD/year).

Postings and pay by role — Volume and median pay by role family.

Skill co-occurrence — Which skills appear together (Python+SQL is the backbone).

🤖 Predicting advertised salary

693 postings carry a genuinely advertised salary, from 293 companies. The interesting part is where the accuracy came from.

Model	MAE	R² (grouped)
Baseline: predict the median	$50.4k	-0.02
Role + skills only	$46.5k	0.12
+ location / source / market	$37.4k	0.37
+ description features	$37.0k	0.38
Gradient boosting	$39.9k	0.27
Blend of two ridge models	$33.5k	0.46

R² went from 0.12 to 0.46 without changing the algorithm. Every point came from data work: normalizing currencies and pay periods (+26% usable rows), using the location, source and category columns that were already collected but never read, and mining the ~6,600-character job descriptions for required years of experience. Gradient boosting lost to the linear model.

Predicted vs actual salary on held-out companies — Out-of-fold predictions on companies the model never saw in training.

A leak I found in my own evaluation. 293 companies supply those 693 postings, so a random train/test split routinely puts the same employer on both sides and the model simply memorises that company's pay band. Switching to GroupKFold by company lowered the headline R² from 0.56 to 0.46. The lower number is the real one, and both are reported side by side.

How it works

Collect → normalize → extract skills → analyze → model → dashboard. Tested (55 unit tests, CI on every push, including a smoke test that runs the whole dashboard) and reproducible.

Sourcing
6 public APIs (Remotive, Jobicy, RemoteOK, Adzuna, Jooble, JSearch) into one normalized, deduplicated warehouse, refreshed weekly by GitHub Actions.

Skill extraction
Curated taxonomy + a matcher that resolves ambiguous names (R vs "R&D", Java vs JavaScript, Excel vs the verb).

Modeling
Baselines first, cross-validation grouped by company, and interpretable coefficients — designed so the honest number is the one on display.

Honest by design: Adzuna-predicted salaries are excluded from pay stats; demand is audited per source (some APIs truncate descriptions); currency conversion uses a dated, static rate table so any chart can be reproduced later; salaries outside a plausible band are counted in an audit rather than silently dropped; and Indonesian postings rarely advertise salary — all stated openly rather than papered over.