Financial inclusion / Methodology

How the inclusion layer is built

Sources

Two files, both long format (iso3, year, indicator code, value). Global Findex Database (CC BY-4.0): 348,889 rows, 162 economies, 5 survey rounds from 2011 to 2024. The Global Financial Development Database: 276,943 rows, ending in 2021. Only the GFDD supply-side indicators are used here (GFDD.AI.02, bank branches per 100,000 adults, and GFDD.AI.25, ATMs per 100,000 adults), both sourced by the GFDD dictionary to the Financial Access Survey (FAS), International Monetary Fund (IMF).

Rounds, not years

The Findex file carries 6 distinct calendar years but 5 survey rounds. The 16 economies dated 2022 have no rows in 2021 and no economy appears in both years, so 2022 is the tail of the 2021 round rather than a round of its own. Grouping by calendar year would report a 16-economy round and drop those economies from every 2021 statistic.

The fold is derived from the data at build time rather than hardcoded: a year continues the previous round when it is within two years of it and shares no economy with it. Both conditions bind, and the table below is what the rule produced on this vintage.

Which economies get a page

An economy gets a page when it has an account reading (account.t.d) in two or more rounds, so that a change is defined. That is 155 of the 162 economies in the file. The other 7 appear in exactly one round and are named on the index instead: Bhutan (2014), Djibouti (2011), Maldives (2017), Puerto Rico (2014), Qatar (2011), Somalia (2014), Syria (2011). The route is generated from that query, so an economy with no data gets no page rather than an empty one.

Coverage is not constant across rounds

Every cross-round figure on these pages runs on a balanced panel, because the reporting set changes between rounds. The table is the number of economies with a value for each headline indicator, by round.

The 2024 round is the sharpest break: 42 economies report account ownership but not the digital-payment indicator, and 39 of them did report it in the 2021 round. The saved and borrowed indicators are restricted to exactly the same set of economies as the digital-payment indicator. The 42 economies with no 2024 digital-payment reading are: Australia, Austria, Bahrain, Belgium, Canada, Chile, Cyprus, Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hong Kong, Hungary, Iceland, Ireland, Israel, Italy, Japan, Kuwait, Latvia, Lithuania, Malta, Netherlands, New Zealand, Norway, Portugal, Russian Federation, Singapore, Slovakia, Slovenia, South Korea, Spain, Sweden, Switzerland, Taiwan, Trinidad and Tobago, United Kingdom, United States, Uruguay.

Why GFDD is not mixed with Findex

The GFDD carries a household-access block that looks like an independent second opinion on Findex and is not one: GFDD.AI.05 (“ Financial institution account”) is Findex re-published on an older vintage. In 2011 it is the same number as the Findex account series: across 143 economies the largest absolute difference is 6.7e-6 percentage points. By 2021 the two differ by up to 38.5 points, because Findex counts a mobile money account as an account and GFDD.AI.05 does not.

That reading is testable, and it holds: the difference between the two series never exceeds Findex’s own mobile-money account share (mobileaccount.t.d). Across the 66 economies carrying all three series in 2021, there are 0 exceptions. The two series measure different things, so this layer never mixes them, and takes from the GFDD only the supply-side counts that Findex does not collect.

The branch series carries impossible values

Saudi Arabia reports 1597.4 bank branches per 100,000 adults in 2020, which is one branch for every 63 adults, and 14,336 ATMs per 100,000 adults, which is one ATM for every 7 adults. The median across the 159 economies reporting a branch count that year is 13.6.

This is not a build error. The value in this file is the value the World Bank publishes: cross-checked on 2026-07-13 against the World Bank API, which returns the same number. It is published here unadjusted, because silently winsorising a source is worse than showing what the source says. It is also why the branch section of the index reports a rank correlation and refuses the level correlation, and why the branch and ATM readings on a country page are labelled as the source’s own.

The GFDD also stops in 2021, and its last year is thin. Branch counts (GFDD.AI.02) are reported for 46 economies in 2021, 159 economies in 2020, 166 economies in 2019, 171 economies in 2018, 173 economies in 2017. A cross-section on the last year would silently drop most of the world, so the branch chart on the index uses the last year with at least 100 reporting economies, derived at build time rather than fixed.

The survey, and what it cannot resolve

Findex is a survey, not a census of accounts. The 2024 round was fielded by Gallup as part of the Gallup World Poll, which interviews about 1,000 randomly selected adults per economy. The World Bank publishes a design effect and a margin of error for each economy’s full national sample: in the 2025 report’s methodology appendix those margins run from 2.1 to 5.4 percentage points, on a median of 1,000 interviews per economy (Global Findex Database 2025, table A.1).

Those margins are for the full sample. Every demographic cut on these pages (women, rural, the poorest 40%) rests on a fraction of that sample, so its margin is wider, and a gap is a difference between two such estimates. The file in this estate carries none of that: its columns are iso3, year, indicator_code, value, with no standard error and no sample size. A gap of a few points between two subgroups, and a change of a few points in a gap between two rounds, cannot be distinguished from zero with this file alone. The counts of economies whose gap narrowed or widened on the index are counts of point estimates moving, and nothing more.

Indicators published

Names and units are the dictionary’s own. The build asserts that every code above still exists in the dictionary and still carries rows in the data file, and throws if one does not, so a renamed indicator stops the build instead of blanking a section.