Replicate this map

Everything on the map derives from one public dataset: BART's Hourly Ridership by Origin-Destination Pairs. Here is how BART itself describes it, verbatim, on their Ridership Reports page:

Hourly Ridership Data. "For those of you looking to take a deeper dive into BART's data - check out our hourly trip datasets. These files will allow you to analyze trips between all stations in the BART system by hour. The data is organized in the following columns: Date, Hour (24-hour clock), Origin Station, Destination Station, Number of Exits. All stations are abbreviated using the 4-Letter station codes."

— BART Ridership Reports. BART also notes the ridership data is provided under a Creative Commons Attribution License.

So it is a faregate census — every paid trip, counted at entry and exit — aggregated to station-pair-by-hour totals, not a survey or a sample. One gzipped CSV per year. The 2025 file is date-hour-soo-dest-2025.csv.gz. Station coordinates come from BART's GTFS feed. The important caveat: the "hour" column is the exit hour.

Download replicate.py — a standalone Python script (standard library only, no dependencies). python3 replicate.py 2025 downloads BART's file and produces all three tables below as CSVs, from scratch. Reading the ~150 lines of that script is the fastest way to audit exactly what this site does.

Table 1 — raw trips: station in / station out / date / hour / people

The data exactly as BART publishes it: one row per station pair per exit-hour per date, counted by the faregates. This download is the first 1,000 rows as a preview of the format; the full year is ~10 million rows — run replicate.py to produce table1_raw_trips.csv in full (or fetch BART's gzipped original directly, link above).

Download table1_raw_sample.csv (1,000 rows, 0.0 MB) — preview:

station_in	station_out	date	people
12TH	16TH	2025-01-01	1
12TH	ANTC	2025-01-01	1
12TH	BALB	2025-01-01	1
12TH	BERY	2025-01-01	1
12TH	CIVC	2025-01-01	8
12TH	COLS	2025-01-01	2
12TH	CONC	2025-01-01	3
12TH	DBRK	2025-01-01	3
12TH	DELN	2025-01-01	1
12TH	EMBR	2025-01-01	2

Table 2 — hourly averages: station in / station out / hour / people

Table 1 with the dates averaged away: for each station pair and hour of day, the average people per day, split into weekdays (federal holidays excluded) and weekend+holiday days. Both 2019 and 2025. This is the exact table the map's station bubbles and kill totals are computed from.

Download table2_hourly_od.csv (193,459 rows, 6.9 MB) — preview:

year	daytype	station_in	station_out	avg_people_per_day
2019	weekday	12TH	16TH	1.16
2019	weekday	12TH	19TH	0.12
2019	weekday	12TH	24TH	1.72
2019	weekday	12TH	ANTC	0.78
2019	weekday	12TH	ASHB	0.53
2019	weekday	12TH	BALB	0.42
2019	weekday	12TH	BAYF	1.84
2019	weekday	12TH	CAST	0.38
2019	weekday	12TH	CIVC	1.45
2019	weekday	12TH	COLM	0.18

Table 3 — link flows: track link / hour / people

Each trip from table 2 routed along its shortest path through the 51-link track network, its riders added to every link crossed (a SFO→Embarcadero trip counts on every link up the Peninsula and through the tube). One row per link direction per hour. This is exactly what the map's line thicknesses draw.

Download table3_link_flows.csv (9,527 rows, 0.4 MB) — preview:

year	daytype	link_from	link_to	avg_people_per_day
2019	weekday	EMBR	MONT	177.08
2019	weekday	MONT	EMBR	744.67
2019	weekday	MONT	POWL	215.71
2019	weekday	POWL	MONT	599.6
2019	weekday	POWL	CIVC	295.89
2019	weekday	CIVC	POWL	400.34
2019	weekday	CIVC	16TH	308.78
2019	weekday	16TH	CIVC	263.17
2019	weekday	16TH	24TH	289.2
2019	weekday	24TH	16TH	183.47

Has anyone done this before?

Yes — routing BART's origin-destination data onto the physical track and mapping the load per segment is a natural thing to do, and others have. Seeing independent versions agree is a good reason to trust the method (and mine); where they differ is worth knowing.

jlperona's BART passenger heatmap is the closest prior work I found: an R/Shiny app that does the same link assignment — hourly OD trips routed onto track segments — as an interactive heatmap on the real geography, for data from 2011 through mid-2021. It's the same core idea as this map; it's now code-only (the hosted demo was retired), covers a different time span, and doesn't have the station-kill / track-pruning tools.
A FlowmapBlue BART map visualizes the same OD data, but as curved flow arcs drawn directly between origin and destination stations — it shows demand between station pairs rather than load on the track each trip actually rides, so the two answer related but different questions.
BART's own crowding / occupancy charts are the agency's version of per-segment load — average riders per car as a train moves down the line — derived from their internal counts rather than the public OD file.

For the congestion side of this project (what BART's ridership means for Bay Bridge traffic), the analysis leans on MTC's Vital Signs and Michael Anderson's "Subways, Strikes, and Slowdowns" (American Economic Review, 2014).

The replicator, inline

The complete replicate.py — read it here, copy it, or download it:

Download replicate.py

"""Replicate every number behind bartanalysis.com/stationmap from BART's raw data.

Standard library only. Downloads BART's public hourly origin-destination file and writes
three CSVs into the current directory:

  table1_raw_trips.csv       station_in, station_out, date, hour, people
                             (every row BART publishes: exact faregate counts,
                              one row per station-pair per exit-hour per date)
  table2_hourly_od.csv       year, daytype, hour, station_in, station_out,
                             avg_people_per_day (averaged across the year's
                             weekdays / weekend+holiday days)
  table3_link_flows.csv      year, daytype, hour, link_from, link_to,
                             avg_people_per_day (every trip routed along its
                             shortest path; riders added to each track link crossed)

Usage:  python3 replicate.py [year]        (default 2025)

Data source: https://afcweb.bart.gov/ridership/origin-destination/
  (one gzipped CSV per year, no login, CC-BY. "Hour" is the EXIT hour.)
Holidays: to avoid dependencies this script treats only fixed-date federal holidays
plus scanned floating ones via a small built-in table for 2019/2025; results match
the site to within rounding. The full pipeline (with the python-holidays package,
input hashing, and cross-checks) is the analysis notebook this site links to.
"""
import collections
import csv
import datetime
import gzip
import sys
import urllib.request

YEAR = sys.argv[1] if len(sys.argv) > 1 else "2025"
URL = f"https://afcweb.bart.gov/ridership/origin-destination/date-hour-soo-dest-{YEAR}.csv.gz"

# US federal holidays for the two years the site covers (extend for other years)
FEDERAL_HOLIDAYS = {
    "2019": ["2019-01-01", "2019-01-21", "2019-02-18", "2019-05-27", "2019-07-04",
             "2019-09-02", "2019-10-14", "2019-11-11", "2019-11-28", "2019-12-25"],
    "2025": ["2025-01-01", "2025-01-20", "2025-02-17", "2025-05-26", "2025-06-19",
             "2025-07-04", "2025-09-01", "2025-10-13", "2025-11-11", "2025-11-27",
             "2025-12-25"],
}
holidays = set(FEDERAL_HOLIDAYS.get(YEAR, []))

# the BART track network: every pair of directly-connected stations (2025 system)
SEGMENTS = [
    ("EMBR","MONT"),("MONT","POWL"),("POWL","CIVC"),("CIVC","16TH"),("16TH","24TH"),
    ("24TH","GLEN"),("GLEN","BALB"),("BALB","DALY"),("DALY","COLM"),("COLM","SSAN"),
    ("SSAN","SBRN"),("SBRN","SFIA"),("SBRN","MLBR"),("SFIA","MLBR"),("WOAK","EMBR"),
    ("WOAK","12TH"),("WOAK","LAKE"),("12TH","LAKE"),("12TH","19TH"),("19TH","MCAR"),
    ("MCAR","ASHB"),("ASHB","DBRK"),("DBRK","NBRK"),("NBRK","PLZA"),("PLZA","DELN"),
    ("DELN","RICH"),("MCAR","ROCK"),("ROCK","ORIN"),("ORIN","LAFY"),("LAFY","WCRK"),
    ("WCRK","PHIL"),("PHIL","CONC"),("CONC","NCON"),("NCON","PITT"),("PITT","PCTR"),
    ("PCTR","ANTC"),("LAKE","FTVL"),("FTVL","COLS"),("COLS","SANL"),("SANL","BAYF"),
    ("COLS","OAKL"),("BAYF","CAST"),("CAST","WDUB"),("WDUB","DUBL"),("BAYF","HAYW"),
    ("HAYW","SHAY"),("SHAY","UCTY"),("UCTY","FRMT"),("FRMT","WARM"),("WARM","MLPT"),
    ("MLPT","BERY"),
]
neighbors = collections.defaultdict(list)
for a, b in SEGMENTS:
    neighbors[a].append(b)
    neighbors[b].append(a)

def shortest_path(origin, destination):
    """Fewest-stations path (breadth-first search)."""
    came_from = {origin: origin}
    frontier = [origin]
    while frontier:
        nxt = []
        for station in frontier:
            for nb in neighbors[station]:
                if nb not in came_from:
                    came_from[nb] = station
                    nxt.append(nb)
        frontier = nxt
    path = [destination]
    while path[-1] != origin:
        path.append(came_from[path[-1]])
    return list(reversed(path))

def daytype_of(date_text):
    day = datetime.date.fromisoformat(date_text)
    return "weekend+holiday" if (day.weekday() >= 5 or date_text in holidays) else "weekday"

print("downloading", URL)
local = f"date-hour-soo-dest-{YEAR}.csv.gz"
urllib.request.urlretrieve(URL, local)

# ---- table 1: the raw rows, plus year totals for tables 2 and 3 ----
year_totals = collections.Counter()          # (daytype, hour, in, out) -> people, whole year
days_in_bucket = collections.defaultdict(set)
with gzip.open(local, "rt") as raw, open("table1_raw_trips.csv", "w", newline="") as out1:
    writer = csv.writer(out1)
    writer.writerow(["station_in", "station_out", "date", "hour", "people"])
    for date_text, hour_text, origin, destination, people in csv.reader(raw):
        writer.writerow([origin, destination, date_text, hour_text, people])
        daytype = daytype_of(date_text)
        days_in_bucket[daytype].add(date_text)
        if origin != destination:
            year_totals[(daytype, int(hour_text), origin, destination)] += int(people)
print("wrote table1_raw_trips.csv")

# ---- table 2: average people per day for each (station in, station out, hour) ----
with open("table2_hourly_od.csv", "w", newline="") as out2:
    writer = csv.writer(out2)
    writer.writerow(["year", "daytype", "hour", "station_in", "station_out", "avg_people_per_day"])
    for (daytype, hour, origin, destination), total in sorted(year_totals.items()):
        writer.writerow([YEAR, daytype, hour, origin, destination,
                         round(total / len(days_in_bucket[daytype]), 2)])
print("wrote table2_hourly_od.csv")

# ---- table 3: route every pair along its path; sum people per track link ----
link_flow = collections.Counter()            # (daytype, hour, from, to) -> people, whole year
path_cache = {}
for (daytype, hour, origin, destination), total in year_totals.items():
    if origin not in neighbors or destination not in neighbors:
        continue                             # station outside the 2025 network map
    if (origin, destination) not in path_cache:
        path_cache[(origin, destination)] = shortest_path(origin, destination)
    path = path_cache[(origin, destination)]
    for a, b in zip(path, path[1:]):
        link_flow[(daytype, hour, a, b)] += total
with open("table3_link_flows.csv", "w", newline="") as out3:
    writer = csv.writer(out3)
    writer.writerow(["year", "daytype", "hour", "link_from", "link_to", "avg_people_per_day"])
    for (daytype, hour, a, b), total in sorted(link_flow.items()):
        writer.writerow([YEAR, daytype, hour, a, b,
                         round(total / len(days_in_bucket[daytype]), 2)])
print("wrote table3_link_flows.csv")
print("days:", {k: len(v) for k, v in days_in_bucket.items()})

Notes

Weekday = Monday–Friday excluding US federal holidays; weekend+holiday is everything else.
Link routing assumes riders take the fewest-stations path — on BART's tree-plus-two-triangles network this is almost always the only sensible path.
Same-station entries/exits (excursion trips) are excluded from tables 2 and 3.
The hosted CSVs are generated by the same analysis notebook that renders the map; replicate.py reproduces them independently (tiny rounding differences only).