Signal & Ledger: Connectivity and Financial Accounts in World Bank Data

In Ghana, 81% of adults have an account at a financial institution or a mobile money provider, and 55% used the internet in the past three months.¹ For a bank or a telco planning an online financial service, these figures don't show how many adults without an account are already online.

The published national figures can't supply that number, because they report account ownership and internet use separately, without showing how often both belong to the same person. Where the data allow only a range, this note reports the range, and it tests each forecast and score on data it wasn't built on.

It draws on three World Bank sources. Global Findex 2025 reports account ownership, internet use and phone ownership among adults aged 15 and over. The World Bank Group Scorecard records the results that development projects report. The World Development Indicators supply population and income per person.

Key findings:

1. Across 139 economies, between 766 million and 1.54 billion adults have an account but didn't use the internet, and between 297 million and 1.07 billion used the internet without an account. These ranges don't include survey sampling error.
2. Of 146 economies, 49 are projected at or above 90% account ownership in 2030, 37 of them already there. For 41 economies the likely range runs both below and above 90%, so the count of 49 could move either way.
3. In the 320 World Bank projects examined, commitments are larger where more adults are unbanked, mostly because larger countries have more of both. Once country size is allowed for, the remaining link could be anything from about zero to moderate.
4. In low- and middle-income economies, men's lead over women in account ownership roughly halved between 2011 and 2024. The narrowing comes from India and China. Without them, the gap widened.
5. A credit-style score ranks project results better than chance, but only modestly. At most, it could help decide which results to review first.

The charts are interactive. Hover for values, switch between measures, or open the Signal & Ledger explorer to see the data together.

1. Internet Use vs Accounts: How many adults have one without the other?

In Ghana, between 5.9 million and 10.0 million adults have an account but didn't use the internet recently. The published shares can't narrow that further.

Take 100 adults: 81 have an account and 55 used the internet. Even if all 55 internet users have accounts, another 26 account holders are left over. At the other extreme, all 45 adults who didn't use the internet could have accounts. So the number lies between 26 and 45 in every 100 adults. These limits are known as Fréchet–Hoeffding bounds, and reporting a range when the data allow no single number is known as partial identification.²

For a service looking for new customers online, the published figures help even less. They are consistent with anywhere from none to all of Ghana's 4.1 million adults without an account having used the internet.

Findex also publishes these shares separately, in most economies, for women and men, two age groups, two income groups, and adults in and out of the labour force. Working out the possible overlap inside each group and adding the groups up, each weighted by its share of adults, rules out some combinations that the national figures alone allow. The weights come from the published shares themselves.³ Across 139 economies, the narrowed ranges are:

Adults, 2024	Lower bound	Upper bound
Have an account but didn't use the internet	766 million	1.54 billion
Used the internet but have no account	297 million	1.07 billion
Own a mobile phone but have no account	685 million	1.21 billion

These bounds show what the published survey estimates allow. They don't include sampling error. Each estimate comes from a national survey sample, typically about 1,000 adults, and Findex doesn't publish standard errors through this source, so the bounds carry no margin of error yet.

In the same 139 economies, 4.53 billion adults have an account and 4.06 billion used the internet. India alone makes up roughly the whole 469 million difference. Without India, the two totals are almost equal.

Account ownership is ahead of internet use in 74 economies and behind it in 65. Regions split more cleanly: internet use is ahead in all 20 Latin American and Caribbean economies surveyed, and account ownership in all 17 surveyed economies in Eastern and Southern Africa.

The World Bank reports about 900 million adults without an account who own a mobile phone.⁴ That figure falls inside the range here, 685 million to 1.21 billion, but it is a weak check: the range is wide, and the two figures may not cover the same economies. The ranges here cover 139 economies with about 95% of the world's adults.

2. Accounts in 2030: Which economies are on course for 90%?

The projections cover 146 economies with at least two surveys, the latest from 2021 or later. The central projections put 49 of them at or above 90% account ownership in 2030, and 37 of those were already there in their latest survey.

Each projection comes with an 80% band, a range meant to contain the 2030 value four times in five. For 41 economies, that band runs both below and above 90%, so the count of 49 could move either way.

Ghana is one of the 41. Account ownership was 81.2% in 2024. The central projection reaches 86.4% in 2030, with an 80% band of 74.2% to 93.4%, so 90% sits inside the band but above the central estimate.

The forecast averages two simpler ones, because neither did well alone. A growth model fitted to the 2011, 2014 and 2017 surveys missed the 2024 shares by 11.3% of adults on average, worse than assuming nothing would change: carrying each economy's last survey forward missed by 10.1%. The equal-weight average of the two had the lowest error in both test years, 5.8% of adults in 2021 and 6.1% in 2024.⁵ We chose the average after seeing both test years, so those errors are somewhat optimistic.

The band's width comes from the 2024 forecast errors, and we tested it on economies not used to set it. With the width set on a random half of the economies, the band held 79.4% of the other half's 2024 values on average across 1,000 random splits, and between 66.7% and 89.4% in nine splits out of ten. The width comes from seven-year-ahead errors. Most projections run six years, from 2024 to 2030, but six economies start from a 2021 or 2022 survey, so their projections run eight or nine years and their bands are likely too narrow.

The growth model is fitted to the economies together on the log-odds scale, a transformation that keeps projected shares between 0% and 100%, and the band has the same width for every economy on that scale. Each economy's trend is pulled toward its region's average trend, so estimates for an economy with few surveys use information from the others in its region.⁶

3. Commitments vs Need: Does World Bank money follow the unbanked?

In the projects examined here, World Bank commitments are larger where more adults are unbanked, mostly because larger countries have more of both.

The analysis covers 320 projects that report broadband, digital-services or financial-services results. Their net commitments total $65.9 billion, counting each project once.⁷ These are commitments, not money already spent, and they fund whole projects, not only the parts aimed at adults without an account. Need is measured in 2024, while the projects were approved between fiscal years 2012 and 2025, three in five of them by 2022.

Across 89 low- and middle-income economies, the rank correlation between commitments and the number of adults without an account is 0.562, on a scale from −1 to 1 where 0 means no link. Holding adult population fixed, it falls to 0.221, with a 95% interval from −0.046 to 0.455,⁸ so the remaining link could be anything from about zero to moderate.

Commitments per adult are also somewhat higher where a larger share of adults has no account (rank correlation 0.246, or 0.346 across 86 economies when need is taken from the 2021 survey). That link mostly reflects differences between the World Bank's low, lower-middle and upper-middle income classes. Within each class, the correlation is between 0.051 and 0.068.

Eleven of the projects cover more than one economy. The ledger doesn't record how their money divides between countries, so we split it equally. Leaving those 11 projects out changes the results little: the first correlation moves from 0.562 to 0.572, and the size-adjusted one from 0.221 to 0.214.

Among low- and middle-income economies with a reporting project, the median commitment is $79 per adult without an account. China, Indonesia, India and Pakistan, the four funded economies with the most unbanked adults, receive between $5 and $28. Commitments per unbanked adult fall as the number of unbanked adults grows, so low figures for these four are what size alone would produce and aren't by themselves evidence of underfunding.

Ghana has two projects here, with combined commitments of $450 million. Between them, they report broadband and digital-services results. Neither reports against the financial-services indicator, so the reported results don't show whether these projects help any of Ghana's 4.1 million adults without an account to open one.

4. The Gender Gap: How far behind men are women on accounts and phones?

Women's account ownership is catching up with men's in the aggregate, but the narrowing comes from India and China. In the other economies taken together, men's lead grew.

In the World Bank's aggregate for low- and middle-income economies, women's account ownership rose from 36.9% in 2011 to 73.0% in 2024, and men's from 46.5% to 77.9%.⁹ Men's lead roughly halved.

Among the 69 low- and middle-income economies surveyed in both years, women's account ownership rose from 38.4% in 2011 to 74.4% in 2024, and men's from 48.8% to 79.3%, so the gap narrowed. Without India and China, women's rose from 30.0% to 57.9% and men's from 38.4% to 67.3%, so the gap widened. It narrowed in only 27 of the 69 economies.

The phone gap and the account gap don't always go together. In Pakistan, women are less likely than men to own either. In Algeria, phone ownership is almost identical, 98.3% among women and 98.4% among men, yet 18.1% of women have an account against 51.9% of men, so phone access doesn't explain Algeria's account gap.

Across the 139 economies with population data, between 352 million and 646 million women own a phone but have no account. For men, the range is 327 million to 573 million. The ranges overlap, so the published shares can't show which group is larger.

5. Project Scoring: Can a credit-style score flag project results likely to miss their targets?

A credit-style score separates stronger and weaker results, but only modestly. Among the highest-scoring fifth, 76.1% reached their target. Among the lowest-scoring fifth, 46.5% did.

At most, that difference could help decide which results to review more closely. The score would be a poor basis for approving or rejecting projects, because close to half of the lowest-scoring results still reached their targets.

Like a bank's application score, the model aims to use only inputs known at approval: the financing arm (IDA or IBRD), lending instrument, responsible sector department, region and approval year, the indicator, its target and the number of indicators the project reports on, and the country's GNI per capita from two years before approval, the latest year then published, in today's revised figures. The ledger's commitment figure and the country's fragility and income flags are left out because they are recorded after approval. Two inputs fall short of the aim: the target and the number of indicators are as recorded at closing, so a restructuring may have changed them after approval.

The approach builds on earlier work that related country and project characteristics to the World Bank's overall project ratings.¹⁰ Here the inputs are limited to approval time as far as the data allow, and the outcome is narrower: whether one results indicator met its target.

The data are 1,130 project-indicator results from 552 closed projects in 115 economies and two regional groupings. A project contributes one result for each indicator or sub-indicator it reports on, and we removed repeat records where the ledger lists a multi-country project more than once. Each result was scored by a model fitted without that result's country, so the test measures performance in countries outside the training data. It doesn't test the score on projects approved after those it was fitted on, which is how it would be used.

The main model, a logistic regression, has a Gini of 0.230, with a 95% interval of 0.116 to 0.334. Gini is the ranking measure used for bank credit scores: 0 means no better than chance and 1 a perfect ranking. A second model, gradient boosting, ranked slightly worse, and we picked the better of the two on the same results reported here, so these figures are slightly optimistic.

Scores are also read as probabilities, so we checked those separately against a flat guess that gives every result the training-data success rate, about 59.5%. The logistic regression's raw probabilities beat that guess only slightly, with a Brier skill score of 0.028, meaning their average squared error is 2.8% lower than the flat guess's. Gradient boosting's probabilities did worse than the flat guess. After recalibration fitted only on the training countries,¹¹ the logistic regression's Brier skill score rises to 0.047. The largest miss in any fifth of the score is then a predicted 60.5% against an observed 55.3%.

These results justify further testing of the score as a way to set review priorities. They don't show that using it would improve project decisions. Information collected during supervision is the next set of inputs to test.

6. What This Is and Isn't

National shares leave the overlap between account ownership and internet use unknown, in section 1 and again for women and men in section 4. Forecasts rest on five surveys at most. Project commitments and reported indicators give only a partial view of who receives support.

The overlap ranges are Fréchet–Hoeffding bounds: the smallest and largest overlap compatible with two published shares. With p the share who used the internet and q the share with an account, both as proportions:

account, no internet use:  max(0, q − p)  to  min(q, 1 − p)
internet use, no account:  max(0, p − q)  to  min(p, 1 − q)

Where Findex reports groups of adults, we work out the bounds inside each group and combine them, weighting each group by its share of adults. Those weights come from the published shares, since a national share is the weighted average of its two group shares. This narrows the ranges, but sampling error remains, and Findex publishes no standard errors through this source to measure it.

We compare the forecasting and scoring models with simple baselines, on later surveys or on countries left out of training. Where the forecast or the score did worse than its baseline, the text says so. The forecast was chosen after seeing both test years, and the scoring model was chosen on the same results it was tested on, so both sets of reported accuracy are somewhat optimistic. The 80% band was tested on economies not used to set its width.

The analysis describes associations and tests predictions. It doesn't establish what causes account ownership to rise or projects to meet their targets. The project scores test a method and don't judge the quality of any project. Account ownership here doesn't measure how often people use their accounts.

Data: World Bank Group Scorecard API, Global Findex Database 2025 and World Development Indicators, downloaded 6 to 14 October 2026. Data credit: World Bank.

References

• Alkema, L., Kantorova, V., Menozzi, C., & Biddlecom, A. (2013). National, regional, and global rates and trends in contraceptive prevalence and unmet need for family planning between 1990 and 2015: A systematic and comprehensive analysis. The Lancet, 381(9878), 1642–1652.
• D'Orazio, M., Di Zio, M., & Scanu, M. (2006). Statistical matching: Theory and practice. Wiley.
• Denizer, C., Kaufmann, D., & Kraay, A. (2013). Good countries or good projects? Macro and micro correlates of World Bank project performance. Journal of Development Economics, 105, 288–302.
• Lessmann, S., Baesens, B., Seow, H.-V., & Thomas, L. C. (2015). Benchmarking state-of-the-art classification algorithms for credit scoring: An update of research. European Journal of Operational Research, 247(1), 124–136.
• Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2020). The M4 Competition: 100,000 time series and 61 forecasting methods. International Journal of Forecasting, 36(1), 54–74.
• Manski, C. F. (2003). Partial identification of probability distributions. Springer.
• Molinari, F. (2020). Microeconometrics with partial identification. In Handbook of Econometrics (Vol. 7A, pp. 355–486). Elsevier.
• Nelsen, R. B. (2006). An introduction to copulas (2nd ed.). Springer.
• Niculescu-Mizil, A., & Caruana, R. (2005). Predicting good probabilities with supervised learning. In Proceedings of the 22nd International Conference on Machine Learning (pp. 625–632). ACM.
• Öhler, H., Negre, M., Smets, L., Massari, R., & Bogetić, Ž. (2019). Putting your money where your mouth is: Geographic targeting of World Bank projects to the bottom 40 percent. PLOS ONE, 14(6), e0218671.
• Susmann, H., Alexander, M., & Alkema, L. (2022). Temporal models for demographic and global health outcomes in multiple populations: Introducing a new framework to review and standardise documentation of model assumptions and facilitate model comparison. International Statistical Review, 90(3), 437–467.
• Vovk, V., Gammerman, A., & Shafer, G. (2005). Algorithmic learning in a random world. Springer.

Citation

Adotey, D. (2026). Signal & Ledger: Connectivity and financial accounts in World Bank data. AIDA Labs Research.