Cafe relocation — does the rent buy enough footfall?

Rent

—

footfall increase needed to offset

—

at the selected place

—

probability this location offsets the rent

Busiest streets

How likely each end-of-month result is

Typical month: — — half of months land above this, half below.

Advanced

Footfall estimate at the selected place

Multiplier

Upper bound

Costs beyond the rent

One-off move cost

Spread over (months)

Other monthly change

Calibrating the map

Busiest place in Ambon, as a multiple of your cafe

How this works — and what would make it wrong

1. What "busyness" means here

There is no footfall data for Ambon, so busyness is inferred from what OpenStreetMap does record. The island is cut into 175 m cells and each cell scores four layers, blended by weight:

Layer	Weight	How the cell scores
Road hierarchy	0.30	Nearest road, weighted by class (trunk 1.0 → service 0.15), decaying over 250 m
Anchor institutions	0.25	Sum over markets, ferry terminals, hospitals, schools, mosques and churches, decaying over 800 m
Commercial clustering	0.20	Sum over shops and small amenities, decaying over 300 m
Junction density	0.15	Count of road vertices within 120 m

Each layer is divided by its own 95th percentile so the weights are comparable, then the blend is divided by its 90th percentile so a typical busy cell reads about 1.0. Cells with almost no road or anchor signal are dropped rather than drawn as zero — which is right over sea, and misleading wherever OSM is simply thin. A population layer is picked up automatically if a BPS density file is present; this build has none.

2. Turning busyness into footfall — the one load-bearing judgement

The index ranks places. Only one thing turns a rank into rupiah:

The exponent decides the verdict, and no Ambon measurement sets it. It is worth stating plainly what it is: a claim about how much busier the busiest place on the island is than your cafe. That is checkable by standing in both places, which no exponent ever is.

Two independent signals in the same OSM extract can be pushed at it. Fit an inhomogeneous Poisson process — intensity proportional to index^elasticity, maximum likelihood over all 15,792 cells — using a predictor built only from road proximity and junction density, so it contains no point of interest of any kind and is exogenous to what is being predicted:

What is being located	Elasticity	Implies busiest cell is
Schools and places of worship (n=189) — sited by population, and near-exhaustively mapped in Indonesia	2.65 ± 0.16	2.3× the cafe
Shops and food outlets (n=192) — sited by trade	4.12 ± 0.23	3.6× the cafe

Face value — the default — puts the busiest cell at 2.37×, just above the population estimate and well below the trade estimate. The old default of 0.5 put it at 1.54×, below both. Three further things point the same way: the road layer is capped at 1.0 by construction, so a trunk road in a village scores what a trunk road beside Pasar Mardika does and no traffic volume enters anywhere; the layer normalisation clips the top tail at 8×; and restated as a rent gradient, 0.5 asserts that premises in the Mardika market district let for about 1.5× what they do in suburban Lateri.

None of that is a footfall measurement. Business density is not pedestrian density, and both estimates carry known biases running in opposite directions: OSM maps shops harder downtown, which inflates them; and under free entry, concentration shows up in the number of outlets rather than in the trade each one gets, which deflates them. Treat the calibration as bracketed, not settled — and note the sharpest version of the same point: if the site is being offered at the going market rate for its location, then in equilibrium the move is roughly break-even by construction, and it only pays if the rent is below market or your costs are below a commercial operator's.

3. Turning footfall into odds

Two numbers meet. The bar is exact arithmetic with no proxy in it — the multiplier at which the move washes its face in an average month:

The odds come from a Monte-Carlo simulation whose draws are baked into this page. Each of the 6,000 draws pairs a plausible month's visitor count — drawn from your own records, with Student-t tails so that having only a few months honestly widens the spread — with a plausible contribution margin. The footfall multiplier is log-normal around the map's estimate, with the "surprised if above" bound set 70% higher, because the proxy is unvalidated. Then

A_i = B_i × exp(sigma × z_i)
net_i = median × A_i − B_i − extra cost

and the headline percentage is the share of draws with net ≥ 0. Every control on this page is an exact recomputation over those same draws, checked at build time against the full model. The map's second shading mode uses the identity that net_i ≥ 0 exactly when median ≥ (B_i + cost) / A_i, so sorting those thresholds once gives every cell's probability by binary search — the colours are the same number as the headline, not an approximation of it.

4. What would make this wrong

The multiplier is the weak link and it is not measured. Nothing in the map has ever been checked against a headcount in Ambon.
Everything pivots on one cell — your cafe's. Its score is a single reading from the same proxy as everywhere else. If it is too high, every multiplier on the map is too low. Sanity-check it against what you know of the street.
OpenStreetMap maps some things far better than others. The anchor layer here is dominated by places of worship and schools, because those are what OSM covers thoroughly in Indonesia. Neither generates much weekday cafe trade, so busy-looking residential areas may be overscored and commercial strips underscored.
Blank areas are not necessarily quiet. Cells with too little mapped signal are dropped. Over sea that is right; where OSM coverage is thin it leaves holes that look like emptiness and aren't.
Street names come from roads, not addresses. A cell takes the name of the nearest named street within 200 m, so a place can be labelled from a road it doesn't front onto.
Steady state only. This assumes the new footfall holds from month one. Your own experience was an opening surge that faded, so early months will flatter the lasting level.
The fix is cheap and it is not more modelling. One afternoon on the shortlisted street with a tally counter beats every layer in this map. Two streets counted would replace the calibration above with a measurement.