score = demand + coverage gap − existing presence

Demand = 0.35·population + 0.25·median household income + 0.40·activity density (supermarkets, shopping centres, schools, universities, hospitals, entertainment, sport, post offices, standing in for foot traffic and spend). Coverage gap = 0.70·distance to nearest CBA branch (capped 15 km) + 0.30·competitor presence (competitors thriving where we are absent = validated demand). Existing presence penalises areas already served. Inputs are z-scored across Greater Sydney, then min–max scaled. Demand uses the population within 5 km of the locality centre, not the locality alone. A branch-viability gate then applies: localities with fewer than 20,000 residents in that 5 km catchment, or fewer than 25 residents per km², are marked below viability and excluded from the commercial ranking, and a ramp discounts catchments under 50,000. The floor comes from the network itself: 95% of localities hosting a CommBank branch today have more than 32,000 residents within 5 km, and the median is 206,000, so 20,000 sits well below anywhere the bank actually operates. Service need is never gated, because community need does not depend on commercial viability.

0.25·seniors + 0.25·disadvantage + 0.15·service anchors + 0.15·cash economy + 0.20·branch distance

Seniors = share of residents 65+. Disadvantage = inverted SEIFA IRSAD (lower advantage = less digital substitution). Service anchors = Centrelink offices (3×) + aged care homes. Cash economy = gaming machines per 1,000 residents. Branch distance = distance to the nearest branch of any brand. This score leaves income out on purpose. It asks where the absence of a physical service point causes the most difficulty.

ABS layers are authoritative. Gaming venues and machine counts come from the official L&G NSW list; net profit is published at LGA level only, so the $/resident figure applies the LGA rate to every locality in it. OSM branch, ATM and POI coverage is community-maintained and incomplete, so production would use the bank's own site data instead. Distances are straight-line rather than drive-time. No confidential banking data is used, and every score weight is a starting assumption to be calibrated with the business owners.

Every fact joins dim_geography[sa2_sk] single-direction 1:*. LGA-grain data (gaming profit) is allocated to SA2 and flagged so the visual can disclose the approximation.

Weight changes create a new dim_scenario row and a new score set, so every published number is reproducible and A/B comparable.

Internal facts sit behind their own security role. The model degrades gracefully: with internal tables absent, every public-data visual still renders.

dim_source carries licence, authority and refresh date, so the data-quality note in the UI is generated rather than written by hand.

Azure Maps visual for points and heat; Shape Map (or Icon Map Pro) for locality (SAL) choropleth using a simplified TopoJSON, because full ASGS geometry is too heavy for 920 polygons at interactive speed.

Distances, z-scores and score components are computed in gold, not DAX. POI facts aggregate to locality × category in a summary table; the point-level table is used only when the map is zoomed in.

Census and SEIFA on release (rare); OSM/POI monthly; L&G NSW quarterly; internal transactions daily with incremental refresh. dim_source.last_refreshed drives the freshness banner.

RLS by region on dim_geography, object-level security on internal facts, workspace separation dev / test / prod with deployment pipelines.

Certified endorsement on the semantic model, measure descriptions and synonyms for Q&A, and a published weights register so no one quotes a score without its scenario.

Drive-time isochrones via ArcGIS or Azure Maps route matrix, then the p-median optimisation run outside Power BI with results landed as a scenario.