Illegal and Licensed Gambling Live on the Same Bank Account

What bank records see that surveys cannot

Almost everything the public debate "knows" about the illegal market rests on indirect estimates: site web traffic, self-reported surveys, focus groups. The Gambling Commission says so itself in presenting its report, published on 28 October 2026: this study changes the raw material. Instead of asking gamblers what they do, it reads what their accounts did — anonymised banking transactions, collected with people's consent, and linked to their survey responses, including their PGSI scores.⁠^[1]

The programme's lead, Steven Murphy, presents the exercise as a first: to his knowledge, no regulator or research team had yet linked observed transactions to individual risk profiles at this scale.⁠^[2] We republish the study here figure by figure, because this is the kind of source a debate can rest on: observed payments, not stories.

The panel: 38,000 people, 157.2 million transactions, eight years of statements

The data was gathered by YouGov from people who explicitly consented to their open banking transactions being used, anonymised, for research; the Commission bought the dataset after anonymisation. The panel covers some 38,000 people from July 2018 to July 2026. In total: 157.2 million transactions, 4.8 million of them with licensed non-lottery gambling companies.⁠^[1]

Two method points matter for everything that follows. Transactions with licensed operators were identified by a classification algorithm developed by Yaspa — regulatory data, test purchasing and machine learning — which labels a transaction as gambling when its probability exceeds 0.5. And each person in the panel is linked, through an anonymised identifier, to information YouGov already held about them, including their PGSI score where one exists. The analysis was run in Python and SQL, with technical support from Warwick Business School.⁠^[1]

Finding illegal gambling in the records: the operator's name is not enough

The first difficulty is documented by the first method's failure. The Commission began by searching for illegal transactions through fuzzy matching of bank-statement descriptions against known illegal operator names: it found nothing. Its interpretation: unlike most legitimate businesses, illegal operators rarely or never include their name in the information they pass to payment providers.⁠^[1]

Two exact methods then worked. Exact matching against descriptions taken from bank statements known to involve illegal operators identified 821 transactions. Exact matching on the merchant-name field, using strongly suspected names, identified a further 2,628, of which 23 had already been found by the previous method. Total: 3,426 unique transactions, made by 179 people — average age 44, 115 of them men — with a PGSI score available for 146.⁠^[1]

The central finding: illegal gambling is played alongside licensed gambling

Among people with repeated detected illegal transactions, 91% also gamble with licensed operators. And the sequence of events contradicts the idea of a switch: the Commission measured, month by month, the licensed activity of the 122 people with at least two detected illegal transactions, around their first illegal payment. It concludes that it does not observe a substantial shift away from the licensed market after that first payment — illegal gambling slots into existing activity, and illegal transactions often arrive in clusters, within short time windows, sometimes interleaved with legal ones.⁠^[1]

Our calculations on this published series: average licensed activity moves from 25.4 transactions a month over the six preceding months to 23.1 over the six following, while detected illegal activity settles at 2.1 transactions a month — about 8% of these people's gambling transactions after the first detection. The licensed decline (−2.3) is the same size as the detected illegal activity that appears (+2.1): our figures alone cannot separate pure addition from partial substitution. What the series does establish is co-use — the same people gamble on both sides — and the Commission, which holds the full individual-level data, concludes there was "no substantial shift" away from the licensed market. That reading is consistent with its earlier Consumer Voice work, where illegal activity already appeared alongside legal activity, in the same people's accounts. Our analysis, below, takes the same series by its totals.⁠^[1]

Risk concentrates: 45% with PGSI scores of 8 or more, against 6%

The second finding rests on the link between transactions and surveys. The PGSI (Problem Gambling Severity Index) classifies gamblers from 0 — no measured risk — to 8 or more, the highest risk category. Among the detected illegal-market users whose score is known — 146 of the 179 people detected — 45% sit at 8 or more. Among the 12,543 licensed gamblers in the panel (excluding lottery-only players), the figure is 6%. The gap is therefore 39 percentage points; the 7.5 ratio between the two shares comes second, because ratios of shares built on small samples are fragile: at n = 146, the 95% margin on that 45% is about ±8 points (our arithmetic). The whole distribution is as if reversed: where licensed gamblers concentrate at 0 (68%), only 26% of those detected on the illegal market score 0.⁠^[1]

Which account the money moves through: 96% current accounts, 4% credit cards

The dataset also records the account type used. Of the 3,426 detected illegal transactions, 3,288 (96%) leave from current accounts, and 138 (4%) from credit card accounts. In people: 172 of the 179 (96%) paid from a current account, and 15 (8%) used a credit card at least once. The two shares overlap: 172 + 15 = 187 against 179 people — about 8 people used both account types; they are not a split between two groups. The Commission records the fact without drawing a conclusion from it: the ordinary banking channel carries the bulk of the detected activity, and the credit card — absent from licensed British gambling — appears only at the margin, but it appears.⁠^[1]

Our analysis: the illegal market runs as a second wallet

Everything above is the Commission's study. What follows is our analysis, and it mixes three things we keep distinct at every step: the Commission's findings, credited to it and cited; our calculations on its published tables, flagged as ours, with their margins; and our interpretations, labelled as such.

A second wallet — our name for the Commission's finding. Its figure: 91% of the people with repeated detected illegal transactions also gamble with licensed operators — 111 people out of 122. Its conclusion, in its own words: illegal gambling is "embedded within existing gambling behaviour rather than triggering a sustained shift away from the licensed market"; its earlier Consumer Voice research already called illegal activity usually supplementary to legal gambling, not a substitution of it. What we add is measured geometry: the 95% confidence interval on that 91% runs from 84.6% to 94.9% (our calculation, Wilson method), and the illegal side rides on the same current accounts as licensed gambling, in the same banking lives, often inside the same short windows of time — where both kinds are observed, the study notes, they often occur within short timeframes. One reading caution: these people were found through their bank traces, so they are banked gamblers by construction — the 91% describes the geometry of the detected group, not a frequency in the population.⁠^[1]

The risk gap climbs in steps. From the two published PGSI distributions we compute the cumulative odds ratios: at a threshold of ≥1 (any measured risk), the odds of sitting above the threshold are 6.2 times higher among people detected on the illegal market than among licensed gamblers; at ≥3 (moderate risk and above), 9.1 times; at ≥8 (the highest band), 12.8 times. The detected group's share exceeds the licensed share at every threshold — dominance across the whole scale — and the excess concentrates at the top: this is not "more risk of every kind", it is above all more severe risk. The limits fit in two lines: the counts are rebuilt from rounded percentages (the licensed shares sum to 99%), and across every integer count consistent with the published percentages, the ≥8 odds ratio runs from about 11.8 to 14.1 — a range to read beside the 95% sampling interval, 9.2 to 17.9. And this is an association measured inside a sample, not an effect: the design does not say which way the link runs.⁠^[1]

Counting the illegal market depends on the method that counts. The Commission's two exact methods barely overlap: 821 transactions found by statement descriptions, 2,628 by merchant names, and only 23 by both — 0.7% of the 3,426 in the union. Statement descriptions alone would have reported 76% fewer transactions; merchant names catch 3.2 times as many as descriptions. Any total of "illegal transactions" is therefore an artefact of the method that produced it. We draw the consequence for our own work: we do not convert that thin overlap into an estimate of hidden volume. The two methods are not independent — they read the same statements — and their overlap counts transactions, not people; a capture–recapture calculation would manufacture a "hidden market" figure the data cannot carry. What the data does carry is the observation itself: two readings of the same statements see two almost disjoint markets.⁠^[1]

The monthly series, reread: replacement at the margin, not piling up. Figure 1 is aligned on the first detected illegal payment: the zeros in the negative months are built in by that alignment, and activity concentrates mechanically around the anchor point. The month +1 peak (4.9 illegal transactions on average) and the fall after it therefore cannot be read as pure behaviour — every event-anchored series draws that shape. What survives the caution is the totals test: add the two series together. Before the first detected payment, the total is 25.4 transactions a month; month +1 rises to 28.6, an anchor month inflated by construction; then, from +2 to +6, the total settles back to 24.6 a month — at the level from before, not above it. After the anchor month, illegal gambling thus looks like it replaces part of licensed gambling rather than durably adding to it — a reading we present as suggestive, not as proof, which reaches, through our own totals, the conclusion the Commission itself draws from this figure: no substantial shift away from the licensed market after the first detected illegal payment.⁠^[1]

A layer added in the sixth month — and one that does not leave. Read the illegal series on its own, month by month. For the six months before the first detected payment it sits at zero — by construction, since the series is aligned on that detection. But licensed activity is already running at full pace: between 23.1 and 28.8 transactions a month, with no established trend. The illegal layer therefore does not appear in a vacuum: it is added, after at least six months of observed play, on top of an installed licensed practice — and it replaces no earlier illegal play, because the statements show none in the observed window. Above all, once it appears it does not fall back to zero: after the anchor month's peak (4.9), it settles at about 1.5 transactions a month — 1.8, then 1.4, 1.9, 1.2 and 1.5 from +2 to +6 — and it is still there in the sixth month, while licensed activity carries on at its usual level. The totals stay where they were (the reading above): the layer persists without swelling the whole. For its part, the Commission takes from this figure that illegal transactions make up a small proportion of total gambling in the months following first engagement, with no substantial shift away from licensed play; the layer's persistence, month after month, is our reading of its published series.⁠^[1] The signal in this series is therefore not only in its totals — it is in the layer's duration: modest in volume beside the licensed base, but laid down to last.

Our interpretation — labelled as such. This pattern — an installed licensed habit, then an illegal layer that is added and persists — is the one the addiction literature predicts. In Blaszczynski and Nower's pathways model, repeated exposure moves the gambler through conditioning, then habituation, then chasing⁠^[3]; its 2022 revision sharpens the gambler subtypes and their routes in⁠^[4]. The DSM-5 counts tolerance among the criteria for gambling disorder: needing to gamble with increasing amounts of money in order to achieve the desired excitement⁠^[5]. And the unlicensed market is where an installed habit meets no friction: no affordability checks, no imposed limits, and no GamStop for anyone who has self-excluded — circumventing protection measures, self-exclusion among them, is one of the motivations the Commission itself recorded in its September 2025 study of the drivers of illegal online gambling⁠^[6]. Two cautions close the reading: "first detected" is a detection floor, not proof of a first-ever payment; and this series carries no individual PGSI trajectory. We are describing a pattern consistent with the escalation literature — a progressive strengthening of the habit — not a demonstration of it.

The credit card: closed on the licensed side, present on the illegal side. Since 14 April 2020, operators licensed in Great Britain may no longer accept credit card payments for gambling — licence condition 6.1.2 of the Commission, e-wallets included.⁠^[7] Yet in the detected illegal activity, 138 transactions (4%) were paid by credit card, by 15 people — 8% of the 179 detected (our 95% interval: about 5% to 13%). The intensity per person is lower there — 9.2 illegal transactions on average per card user, against 19.1 per current-account user: the channel is broad but shallow. Our reading: the ban did not eliminate card spending on gambling; it closed the channel where it was visible and regulated, and the spending reappears, small but real, on the illegal side, outside the rule's reach. The volume is not there: it stays on current accounts.⁠^[1]

GamStop stops at the licensing border. Since 31 March 2020, taking part in GamStop — the national self-exclusion register — has been a condition of the British licence: one registration blocks access to every licensed online operator.⁠^[8] The illegal operators in this study hold no British licence and are not in the scheme. For a person in co-use — the case for 91% of the repeat group — self-excluding therefore closes the licensed side of their gambling, while the illegal side, fed by the same accounts, stays outside the block. Protection tools are built market by market; co-use crosses markets.

Why surveys cannot see this geometry. A survey asks people what they do, market by market. The co-use the statements show has exactly the shape self-reporting misses: two markets on one account, illegal payments in short clusters, and operators the gambler may themselves believe are licensed. The banking method sees the geometry because it reads payments, not declarations — and that, in our view, is this study's durable contribution: it moves the question from "how many people gamble illegally" to "how illegal gambling fits inside gambling as a whole".⁠^[1]

What this study does not allow anyone to conclude

The Commission publishes its limitations alongside its results; we republish them with the same prominence, because they are part of the result.⁠^[1]

• The sample is not nationally representative. Men are slightly over-represented, women under-represented — especially over 55 — and the panel includes only people who agreed to share their banking data for a small incentive. The PGSI comparison therefore holds inside this sample, and cannot be set against the national survey (GSGB) figures.⁠^[1]
• Part of the market stays invisible. Only cash transactions through open-banking-supporting institutions are captured: cryptocurrency payments to unlicensed operators escape the net. And only a subset of illegal operators is covered — some no longer operating in Great Britain. Undetected illegal transactions therefore remain in the data, including among people already detected.⁠^[1]
• The size of the illegal market cannot be estimated from this data, in absolute terms or relative to licensed gambling, where the vast majority of transactions can be identified.⁠^[1]
• No causality is established. The design cannot say whether illegal gambling worsens risk scores, whether the highest-risk gamblers are drawn to the illegal market, or both.⁠^[1]
• The method discovers no new operators. It recognises payments to operators already identified by other means; on its own it cannot reveal an unknown operator.⁠^[1]

Method and data: the figures, figure by figure

This article's single source: the Gambling Commission's report, Illegal gambling: findings from an open banking data analysis, published on 28 October 2026 (introduction, results, materials and methods, limitations)⁠^[1], together with the presentation post by its scientific lead, Steven Murphy.⁠^[2] No figure comes from press coverage; the only numbers that are not the Commission's are the averages and ratios we computed from its published tables, and they are flagged as ours wherever they appear. In the CSV files below, a source_flag column separates the Commission's figures ("UKGC") from our calculations ("our arithmetic"); the monthly file's total column is also our calculation.⁠^[9]

Two reading cautions for Figure 1, which we add to the Commission's own. Before the first detection, the monthly licensed values move between 23.1 and 28.8 with no established trend: a slope fitted to those six points (−0.7 transactions a month) is statistically indistinguishable from a flat series — drift inside wide month-to-month noise, not a demonstrated decline, and we do not read the pre-period as a fall. The series is also aligned on the first detected illegal payment: the zeros before the anchor are built in by that alignment, activity concentrates mechanically around the anchor point, and the people who appear in the figure are the ones the detection method caught — detection selects who appears in the figure.⁠^[1]

We republish the study's figures here in usable form: the key-figures file — panel, detection, PGSI distribution, account types, each tied to the report section it comes from — and the Figure 1 monthly-series file. This piece extends our editorial "We Measure What Others Narrate", whose first announced item it was; all our studies, with their data, are catalogued at Published Studies.⁠^[9]

• The study's key figures (CSV) — panel, detection, PGSI, accounts.
• The Figure 1 monthly series (CSV) — licensed and illegal activity around the first detected payment.

If you are a journalist or a researcher and you want an extract, a method clarification or a comment: contact the newsroom via the Press page.

Frequently asked questions

References

1. Gambling Commission — Illegal gambling: findings from an open banking data analysis, technical report, 28 October 2026. gamblingcommission.gov.uk ↑
2. Gambling Commission — Steven Murphy, "Exploring the potential of banking transaction analysis to understand consumers' use of illegal gambling", blog post, 28 October 2026. gamblingcommission.gov.uk ↑
3. Blaszczynski, A. & Nower, L. (2002). "A pathways model of problem and pathological gambling". Addiction, 97(5), 487–499. doi.org ↑
4. Nower, L., Blaszczynski, A. & Anthony, W. L. (2022). "Clarifying gambling subtypes: the revised pathways model of problem gambling". Addiction, 117(7), 2000–2008. doi.org ↑
5. American Psychiatric Association (2013). Diagnostic and Statistical Manual of Mental Disorders (DSM-5). American Psychiatric Publishing — gambling disorder, tolerance criterion. ↑
6. Gambling Commission — Illegal online gambling: consumer awareness, drivers and motivations, research report (with Yonder Consulting), 18 September 2025. gamblingcommission.gov.uk ↑
7. Gambling Commission — "Preventing credit card use", licensee guidance: licence condition 6.1.2, in force since 14 April 2020, bars licensed operators from accepting credit card payments for gambling, including through a money service business. gamblingcommission.gov.uk; and "Gambling on credit cards to be banned from April 2020", news article, 14 January 2020. gamblingcommission.gov.uk ↑
8. GamStop — Terms of Use: every business licensed to offer online gambling in Great Britain must take part in the national self-exclusion scheme, as a condition of its licence, since 31 March 2020. gamstop.co.uk ↑
9. Chicken Road Productions — this article's data (CSV, free to download, with a source_flag column separating the Commission's figures from our calculations): the study's key figures · the Figure 1 monthly series. ↑