Busan Station Dwell Conversion Analysis
Estimated dwell, a variable absent from the source data, using only Busan metro tap-in and tap-out records, and identified stations where arrivals are high but people do not stay.
Period · August 2026 (in progress)
Role · Individual project · entry in preparation for the 2026 Big Data competition
Source · GitHub repository
Quick read
The through-line of the work.
Open
The starting question
Policy picks target sites by boardings and alightings, but alighting means someone arrived, not that they stayed. The problem was that dwell is a variable no public statistic records.
Build
What I checked
Subtract cumulative boardings from cumulative alightings and you are left with the people who came in by rail and have not yet left. I converted the area under that curve into a mean dwell time with Little's Law.
Turn
What changed my view
The first version of the metric collapsed at residential stations: 86.7% of station-days were negative, and the mean dwell time came out at an impossible −1.23 hours. Before doubting the data, I re-examined the reference point I had chosen.
Close
Where I draw the line
Shifting the curve to each day's minimum brought the values back into a normal range, and the correction also showed that the order of the daily trough and peak alone separates inflow stations from residential ones.
01 · Start
The question I began with
What does a ridership-based policy target list miss? Are alighting and dwelling the same thing?
02 · Material
The data and evidence
03 · Work
How I worked through it
- 01
Test the premise first
Compared system-wide boardings and alightings to confirm the data records gate crossings, so inter-line transfers never appear as a boarding or alighting.
- 02
Design the dwell metric
Built net rail-attributable dwelling population as cumulative alightings minus cumulative boardings, then divided the curve area by daily alightings to derive mean dwell time.
- 03
Find and fix the defect
Found the raw metric collapsing to negative values at residential stations and applied a per-day min-shift so only the rail-attributable increment remained.
- 04
Test the competing explanation
Checked whether hotel density explains Haeundae's low dwell using a checkout index, and recorded the conclusion that it does not.
04 · Takeaway
What the numbers suggested
After the min-shift correction, mean dwell time at non-transfer stations moved from an impossible −1.23 hours to 2.78 hours.
The order of the daily trough and peak alone separated 49 inflow-type stations from 63 residential-type stations.
Haeundae station averages 0.96 hours of dwell, the lowest of the 49 inflow-type stations against a median of 2.09, and shows the widest gap between alighting rank and dwell rank.
The lodging explanation was detectable but insufficient: the Sunday checkout index reached 2.72, yet the signal accounts for only 5.8% of inflow and cannot explain the low conversion rate.
Not claimed
- Does not claim dwell time was measured. It is a proxy estimated from gate records.
- Does not claim a verified link to commercial revenue. Merging retail data is still outstanding.
- Does not claim final results. This is first-pass validation ahead of a September 2026 submission.
05 · Use
Where this helps
Ranking stations by failed dwell conversion instead of raw ridership produces a different target list for commercial-district renewal budgets, and producing that list is the point of the analysis.
06 · Boundary
What still needs checking
- •The project is in progress ahead of a September 2026 submission, and current figures are first-round validation results.
- •Commercial-establishment data and regression residual diagnostics are still outstanding, so stations undersupplied relative to dwell are not yet identified.
- •The dwell measure is a proxy for rail-attributable presence and excludes arrivals by bus or car.
- •Only 112 of 114 stations appear in the ridership data, so two stations are missing from the analysis.
07 · Figures
Tables and figures
Lowest dwell among inflow-type stations
Verified sourceRecomputed from the KPI table after the min-shift correction. Haeundae is the lowest of 49 inflow-type stations, under half the median. Dwell is a proxy estimated from gate records.
Source · data/processed/station_kpi.csv · 112 stations, Jan–Jun 2026
Sunday checkout index
Verified sourceHow much boarding concentrates late in the morning on Sundays, when commuting disappears. Haeundae ranks second behind the KTX interchange, so the lodging signal is real — but it covers only 5.8% of inflow and was rejected as the explanation for low dwell conversion.
Source · data/processed/station_typology.csv · Sunday 10–13h ÷ 07–10h boardings