The graduate premium has a long way to fall yet

For a long time now, the graduate premium has been a key aspect of the marketing of higher education.

It was pretty much relied upon to justify the switch to 9k fees – and Rachel Reeves is still using it to justify her repayment threshold freeze at last November’s budget.

It’s been under pressure from all sorts of angles. The rise in the minimum wage erodes it from below, and Longitudinal Education Outcomes (LEO) data erodes it from a provider type, subject and geographical perspective.

But the premium – or at least graduate earnings in general – matters more than for messaging purposes.

Because student loans have to be split up into an asset (the total value of what the government thinks it will get back) and a write-off (the chunk it thinks it won’t, ie the public subsidy), as well as estimating long-term economic factors like growth and inflation (and therefore the value of money over time), an estimation of what those graduates will earn has to be done.

The loan value sits on the balance sheet. The write-off is counted as “current expenditure”. So when the former goes down in value, a whopping extra write-off charge gets added, in-year, to the accounts.

That has happened, this year, in all three devolved nations.

What am I now?

England moved to the new model in 2021-22. That’s when the residual fair value movement – the pure earnings-outlook-driven revaluation – jumped from -£410m in 2020-21 to -£6,304m in 2021-22, then -£6,679m in 2022-23.

But it’s taken a while to apply it elsewhere.

In the supporting documents for the Scottish Spring Budget Revision 2025-26, there’s a whopping a £2.9bn increase in the budget provision for the non-cash revaluation of student loans within the Education and Skills portfolio.

The document says this:

…follows the implementation of a new, more accurate, model for forecasting student loan repayments and changes to the discount factor applied for that valuation.”

In the Northern Ireland Spring Supplementary Estimates 2025-26, there’s a £1.7bn increase in the Department for the Economy’s ringfenced RDEL budget provision for the non-cash revaluation of student loans – with a further £1.1bn added at the Supplementary Estimates stage by Treasury.

The estimates memorandum says the increase is:

…due to the adoption of a new model for estimating the future value of: income from student loans; and residual loans that are not expected to be paid in full at the end of the loan period.

It adds that:

…the updated model better reflects several variables that are used to forecast the future values of the loans, for example borrowers’ future earnings…

and that the change:

…has resulted in a one-off charge in the current year to revalue estimated future payments from existing loans prior to 2025/26 on the new basis, as well as a higher estimated write down for the future value of loans issued in 2025/26 using the new parameters in the new model.

Wales too. In the Welsh Government’s 2nd Supplementary Budget 2025-26, there’s a similarly whopping £2.6bn increase in the budget provision for the non-cash revaluation of student loans within the Student Loans Resource line within the Education department.

The explanatory note says this:

…follows a change to the student loans valuation model which brings the model in Wales in line with the model used by the UK Government’s Department for Education.

When he was asked at the Senedd Finance Committee why the £2.6bn isn’t being given to schools, Cabinet Secretary Mark Drakeford explained that “it’s not that sort of money” – it’s a non-cash allocation that the Welsh Government can’t choose to spend on other purposes.

Mike Hedges MS underlined the importance of making it “absolutely clear” to stakeholders that “this is not money that’s being stolen off schools – it is something entirely different.” Well, it’s not being stolen off schools for now, Mike. In the longer term, however…

We knew this was coming. Avid readers of my “every day’s a school day” journey through student loans accounting in recent months will recall that in January we noted that Audit Wales had qualified the Welsh Government’s accounts, that Audit Scotland had flagged systematic over-forecasting, and that when a new model for predicting earnings used by England’s Department for Education (DFE) was adopted across all the devolved nations, the system was going to look a lot more expensive. The bills have now arrived.

Audit Scotland’s annual audit report for the Scottish Government said the old model’s methodologies were:

… believed to over-forecast student loan repayments, leading to understated impairment charges.

And alarmingly, Audit Wales, in qualifying its opinion on the Welsh Government’s 2024-25 accounts, said:

Actual repayments are currently around 50% of those currently forecast under the existing model.

Mirror stops lying

To work out the value of the loan v the write off, the government builds a computer model – a kind of giant prediction machine – that tries to guess what millions of graduates will earn over the next 30+ years.

In the olden days, this was done via something called the “Student Loan Repayment Model” (SLRM) – a cohort-based micro-simulation used by the Department for Education (DfE).

It tried to guess how much graduates would earn and repay, but it got too many of those guesses wrong, didn’t cope well with the new bigger loan system, and small changes could wildly change the answer, so they replaced it with something more reliable.

In 2010 the government hired the consultancy firm Deloitte to build a replacement called the HERO model, which launched in June 2011.

It covered English student loans – that is, loans taken out by students who normally live in England, even if they went to university in Scotland, Wales or Northern Ireland. It also covered EU students studying at English universities.

But to predict what those English borrowers would earn, it didn’t use data about English graduates specifically. Instead it used a survey called the British Household Panel Survey (BHPS), which tracked roughly 10,000 people across Great Britain (that’s England, Scotland and Wales — but not Northern Ireland) from 1991 to 2008.

The BHPS wasn’t designed for student loan modelling – it was a general survey of the population – but it was the best source available at the time for tracking how people’s earnings change from one year to the next over long periods.

To predict what a graduate would earn, it used what’s technically called an “income percentile transition matrix”.

Imagine lining up every graduate in the country by their earnings, from lowest to highest, and giving each one a number from 1 to 100 based on where they stand (so the person exactly in the middle gets 50).

The model asked – if you’re at position 40 this year, what are the chances you’ll be at 35, 40, or 45 next year? It did this separately for men and women, and for different age groups.

This is called a “first-order Markov process,” which is a fancy way of saying – the model only looks at where you are right now to predict where you’ll be next. It doesn’t remember where you were two or three years ago.

Once it had your predicted position in the queue, it converted that into an actual salary figure using data from the Labour Force Survey (LFS), a UK-wide survey, which tells you what the person at position 40 actually earns in pounds.

In the dark

A couple of years in, the NAO published a detailed evaluation paper, and their findings were damning. The model still consistently over-predicted repayments – forecasts were 7–9 per cent higher than what graduates actually paid back. Several problems stood out.

First, the BHPS transition matrix was based on survey data from 1991 to 2008 – a period before the financial crisis. The model assumed graduates in the future would move up the earnings ladder at the same rate as graduates in the boom years of the 1990s and 2000s.

But evidence was mounting that newer graduates were earning less, and that the “graduate premium” was shrinking, especially at the bottom.

Second, the model only looked at one year of past earnings to predict the next year. The NAO showed that using five years of past earnings made predictions much more accurate – the error dropped from roughly 17–22 per cent to about 4–8 per cent.

Third, and perhaps most importantly, the model ignored what subject you studied and which university you went to. The NAO ran their own analysis of 2.6 million borrower records from the Student Loans Company and found that these factors made a real, measurable difference to earnings.

A Russell Group graduate earned on average £2,080 more per year than an otherwise identical University Alliance graduate. An Art & Design graduate earned £1,200 less than a Social Sciences graduate. A Law graduate earned £1,380 more. Over a 30-year career, these differences really add up. And the model was blind to all of it.

The NAO also showed something striking about career trajectories – two 22-year-old male graduates, one studying Law, one studying Art & Design, might start at very similar points on the earnings ladder. But five years later, they’d be at opposite ends of the earnings spectrum. The HERO model treated them identically. Deloitte was not asked to hand back its consultancy fee.

Without saying

The good news is that the government was already building a replacement called STEP – the Stochastic Earnings Pathways model (“stochastic” just means it includes an element of randomness, rather than predicting a single fixed outcome for everyone).

STEP addressed the NAO’s two biggest criticisms. It was no longer a “Markov” model that only looked one year back. Instead, it used mathematical formulas (called “regression equations”) that took into account up to five years of previous earnings.

This meant that if you’d been steadily earning £30,000 for the last five years, the model would predict something very different than if you’d just jumped from £15,000 to £30,000 – which makes intuitive sense.

The model also incorporated subject of study and mission group (Russell Group, University Alliance, Million+, GuildHE, etc.) into its predictions. This was a direct response to the NAO’s finding that these factors mattered.

But subject and university were only used for the first three years after graduation, when earnings were predicted using formulas built from Student Loans Company (SLC) records covering 1998 to 2014.

From year four onwards, earnings were predicted still using formulas built from BHPS survey data – and the BHPS didn’t have enough detail to keep tracking subject and university effects.

So from year four the model still assumed that two graduates with the same earnings, age, and gender would follow similar paths regardless of whether one studied medicine at Cambridge and the other studied drama at a small college.

The model worked in two stages. For the first three years after a borrower’s Statutory Repayment Due Date (SRDD – usually the April after graduating), earnings were forecast using formulas built from SLC data.

These predicted earnings based on the borrower’s characteristics – which university group they attended, what subject they studied, their course level, age, whether they were English or EU, their gender, and whatever prior earnings were available.

Different formulas were used depending on gender and whether the borrower had low (under £10,000), medium (£10,000–£30,000), or high (over £30,000) prior earnings. This meant the model could capture the fact that a woman with low prior-year earnings faces different prospects than a man with high prior-year earnings.

From year four onwards, the model switched to a different set of formulas built from the BHPS – still the same Great Britain survey (1991–2009) that HERO had used, though with the data extended by one year.

These used three years of prior earnings, age, and gender. But the BHPS and LFS don’t always agree on what people earn – the surveys are designed differently. So the model didn’t just trust the BHPS salary figures directly.

Instead it used the BHPS to work out where in the queue a graduate would be (e.g. “this person is in the middle of the pack for a 25-year-old woman”), and then looked up what the LFS said the middle-of-the-pack 25-year-old woman actually earns. This ensured the final salary figures were anchored to the more reliable, UK-wide LFS data.

The model also separately predicted whether someone would be employed, unemployed, emigrated, or “other zero” (inactive for some other reason) before estimating their earnings. Employment predictions used a statistical technique called logistic regression – essentially a formula that calculates the probability of something being true or false (in this case: “is this person employed?”) based on their characteristics.

A bit of randomness was also thrown in, so that not every borrower with an 88 per cent chance of being employed would actually be predicted as employed – some would randomly be assigned to the unemployed group, just as happens in real life.

The earnings model was then further split into four separate sub-models depending on the type of qualification the borrower was studying for – a sub-degree course (like foundation degrees), full bachelor’s degrees, teaching qualifications (PGCEs), and people who dropped out before finishing.

Each sub-model used a different comparison group from the BHPS – for example, the dropout sub-model compared borrowers against people in the survey whose highest qualification was A levels, since that’s roughly where a dropout ends up.

The STEP model also modelled several factors beyond basic earnings – an “age at graduation” adjustment, investment income above £2,500, mortality (using a blend of ONS life tables (for England and Wales) and SLC data (the SLC data showing lower death rates than the ONS tables, reflecting that graduates tend to live longer)).

And “repayment frictions” were accounted for – this is the term for the gap between what borrowers should repay based on their annual salary and what they actually repay. For instance, someone who earns most of their money in a few months might overpay through the tax system, while someone with two part-time jobs might underpay. The model adjusted some borrowers’ effective earnings up or down based on these patterns in historical SLC data.

Postgraduate earnings were handled simply – the model took the predicted earnings for a first-degree graduate with the same characteristics and added a flat percentage uplift – 8.9 per cent for male Master’s students, 10.3 per cent for female Master’s students, 8.0 per cent for male Doctoral students, 6.0 per cent for female Doctoral students. The figures came from a single piece of research (Conlon & Patrignani, 2011).

Every piece of me

Alas, the model’s forecast accuracy was still a problem – the model tended to over-predict repayments.

So they tried again. In 2019/20 Longitudinal Education Outcomes (LEO) data was incorporated into the model – linking higher education records from English universities to UK-wide HMRC tax records, creating a dataset that shows what graduates actually earned after university.

That solved a specific problem – SLC data didn’t include the earnings of people who’ve fully repaid their loans. Since the highest earners repay fastest, this means the SLC dataset gradually loses its richest graduates and paints an increasingly gloomy picture of graduate earnings.

By linking SLC borrowers to their LEO records – using approximate matching on names, dates of birth and other details (known as “fuzzy matching” because the details don’t have to be a perfect match) – the model could now track what those high earners were actually earning after they disappeared from the SLC dataset.

In the STEP model, SLC-based predictions covered only the first three years after graduation, and BHPS-based predictions took over from year four. With LEO data available, DfE could see actual earnings for borrowers up to 10 years after graduation.

So they restructured the model into a “short-term earnings model” covering years 1–10 and a “long-term earnings model” covering years 11–43. This was a big deal – the first 10 years of a graduate’s career are when earnings are growing most steeply and when prediction accuracy matters most for getting repayment forecasts right. Subject of study and university group – which had previously been dropped after year three – now influenced predictions for a full decade.

Meanwhile, while the STEP model used mathematical formulas that tried to summarise the relationship between a person’s characteristics and their earnings into a single equation, the 2019/20 model replaced this with a “nearest-neighbour” approach.

Instead of plugging numbers into a formula, the model searched a training dataset (built from SLC and LEO data, covering 2014 to 2017) to find the most similar real person – or group of people – based on variables including university type, subject, whether the student was English or EU, gender, age, whether they dropped out, and up to three years of previous earnings.

To measure how “similar” two people are, the model treated each of these characteristics as a dimension and calculated the distance between the two people in that multi-dimensional space (this is called “Euclidean distance” – think of it as a ruler that works in many dimensions at once). It then used that real person’s actual next-year earnings as its prediction. Where multiple individuals matched, the model selected one at random.

That was a fundamentally different philosophy. Instead of deriving a formula that summarised everyone’s experience, the model said:

Let’s find someone who actually looks like you and see what happened to them.

It’s like the difference between being told “the average footballer earns X” and being shown five specific footballers who play the same position, at the same level, at the same age as you – and seeing what they actually earned.

Begging for footnotes

You know what’s coming. It was somehow still overestimating repayments.

So in 2020/21 the BHPS was dropped entirely, and HMRC administrative earnings data was brought in to replace it.

The long-term model (years 11–43 after graduation) now draws on HMRC administrative earnings data covering 10 per cent of the entire UK population from 2012 to 2020. Instead of a GB-only survey of roughly 10,000 people that stopped collecting data in 2009, the model uses actual tax records for millions of people across the whole of the UK, covering much more recent years.

The HMRC data doesn’t identify whether someone is a graduate or not. DfE’s rationale for using it is that by 11+ years after graduation, simply having a degree doesn’t give you much of an edge any more – what matters is your track record of actual earnings.

So at that career stage, knowing someone’s age, gender, and last three years of earnings is enough to predict their next year – regardless of whether they went to university. This is a debatable assumption – you could argue that a degree still shapes your career options decades later – but it allows DfE to draw on a vastly larger and more current dataset than the BHPS ever provided.

The nearest-neighbour approach that had been introduced for the short-term model in 2019/20 was now applied to the long-term model as well. Both parts of the model now work the same way – find the 10 most similar real people in the (UK-wide) training data, and pick one of their outcomes as your prediction.

People who are closer matches are much more likely to be chosen – specifically, the model weights them so that someone who’s twice as close is four times as likely to be picked.

The voluntary repayments model was upgraded so it could now distinguish between borrowers who pay off their entire remaining balance in one go and those who make a smaller extra payment – previously it couldn’t tell the difference.

Repayment frictions and overseas repayments were modelled using “decision trees” – a technique where the model asks a series of yes/no questions about a borrower (like “are they earning over £30,000?”, “did they make a voluntary repayment last year?”) and follows different branches depending on the answers, like a flow chart. These are incremental improvements rather than fundamental architecture changes.

And it all caused a massive writing down of the loan value on the balance sheet and corresponding hit to the taxpayer of circa £20bn. The only reason we didn’t notice in England was because inflation was running so high, offsetting via more loverly interest being added to the balances of Plan 2 grads.

I’m lost

If you’re still reading, thanks. You might be thinking “shouldn’t applicants know this” – but just imagine. The government has built a model that knows – down to the level of your subject, your university group, your gender, and whether you dropped out – roughly what you’re going to earn and how much of your loan you’ll never pay back. It just doesn’t tell you any of that before you sign up. Many would argue that it should.

If on the other hand you were a government seeking to limit or at least control costs of the system to the taxpayer, you might identify the courses and providers where the model predicts the highest write-offs and do something about them – cap student numbers, withdraw loan funding, that sort of thing. It hasn’t done that either.

Instead, as I’ve been pointing out extensively on the site in recent months, the Treasury has been limiting its exposure in other ways – by holding down with the thresholds for stepped interest and repayment for that big wedge of Plan 2 debt, first in 2022 and now at last November’s budget.

(By the way. The mission group thing. Every version of the published methodology, from the 2017/18 technical notes through to the current publication accompanying the April 2025 forecasts, lists the same seven provider group categories by name– Russell Group, 1994 Group, University Alliance, MillionPlus, GuildHE, Large non-affiliated and Small non-affiliated.

Each category comes with named example institutions. Those examples have never been updated either – “Leeds Metropolitan” is still listed under MillionPlus despite having rebranded as Leeds Beckett in 2014, and the 1994 Group is still listed as a category despite having dissolved in November 2013. These are frozen legacy categories that have not been revisited to reflect how the sector currently organises itself, which institutions have moved between groups, or — critically — how those institutions now deliver their provision.)

But if you think that’s the end of it, well. Strap in.

Someone I don’t want around

As it stands, the model matches borrowers based on provider group (Russell Group, University Alliance, etc.), subject, gender, age, domicile, dropout status, and prior earnings.

But the “provider” is the registered institution – the university that holds the registration and issues the degree.

If a University Alliance (as it was ages ago) university franchises 10,000 business studies places to a private delivery partner, those students appear in the SLC and LEO data as University Alliance business students.

The model has no variable for whether a course was franchised, who actually delivered the teaching, or where it was physically taught. It can’t see the difference between a business student taught on campus at the parent university and one taught by a subcontractor in rented office space.

Then think about timing. The current model trains on SLC data covering 2001–2020 and LEO data covering 2014–2020. The nearest-neighbour matching for early-career earnings draws on a training window from 2014–2020. The big expansion in franchised provision has mostly happened from around 2020 onwards, so the bulk of those students would have SRDDs from roughly 2023–2027. That means they’re barely in the training data yet, if at all.

The model is currently predicting their earnings based on the outcomes of earlier cohorts at the same provider groups studying the same subjects – cohorts that were overwhelmingly directly taught.

So right now the model is almost certainly over-predicting the earnings of those franchised business students, because it’s matching them to the historical outcomes of non-franchised students at the same universities in the same subjects.

As those 100k a year ish franchised students start passing their SRDDs and their actual earnings flow into the SLC and LEO data, two things happen.

First, the training dataset gets populated with a large new group of, say, University Alliance business graduates whose outcomes are substantially worse than the existing University Alliance business graduates in the data. The nearest-neighbour model will start finding these lower-earning individuals as matches for future borrowers. That will drag down the predicted earnings for all students at those provider groups in those subjects – including the ones being taught on campus who may have perfectly fine outcomes. The model can’t tell them apart.

Second, and more importantly for the headline numbers – the taxpayer cost goes up. If the model starts predicting lower earnings for a large and growing chunk of the borrower population, the expected repayments go down, the expected write-off goes up, and the cost of the system to the taxpayer increases. Given that franchised business studies is now one of the largest single pipelines of new borrowers, the impact on the aggregate RAB charge could be substantial.

Falling

For the devolved nations, things get very interesting, and potentially very unfair. A Welsh university – say Cardiff Metropolitan, which sits in the University Alliance group (although isn’t actually a member here in 2026) – teaches a mix of students. Some are Welsh-domiciled (funded by Student Finance Wales), some are English-domiciled (funded by Student Finance England), plus smaller numbers from Scotland, NI and overseas.

When the early-career model looks for nearest neighbours for, say, a 21-year-old woman studying business at Cardiff Met, it matches on provider group (University Alliance), subject (business), gender, age, domicile (England), and prior earnings.

It doesn’t match on which specific university, or which country that university is in. Cardiff Met goes into the same “University Alliance + business studies” bucket as every other University Alliance institution – including the ones in England that are massively expanding franchised business provision.

So as those franchised students’ poor earnings flow into the training data, the model starts finding them as nearest neighbours for the directly-taught Cardiff Met student. The model predicts lower earnings, lower repayments, a higher RAB charge – and Cardiff Met hasn’t franchised anything. Its English students just got tarred because they share a mission group label and a subject with a completely different kind of provision at a completely different institution in another country.

That has a concrete consequence – DfE’s model now says English students at Cardiff Met doing business studies are a worse bet for the taxpayer than they actually are.

Actually, to be fair, I don’t actually know whether “adopted the model” means Wales, Scotland NI are using DfE’s UK-wide training data, or whether it means they are running the same methodology but trained on their own borrower data. HMT won’t tell me.

For the nations, HMT operates three separate spending controls – one on loan outlay, one on all the stuff in this blog, and one on fair value revaluations (ie the value of money in the future, inflation etc).

For that first one, it could be that only Welsh, Scottish and NI grads are being looked at for each calc, and it could be that UK-wide grads are. As I say, HMT won’t tell me. And for all we know, either method could end up making those countries’ grads a better or worse bet, impact their graduate earnings predictions and/or notional graduate premium, and affect what they’re allowed to lend to students on the basis of “equivalence”.

But what I’m fairly sure about is three things. First, that big wedge of franchised business provision isn’t in the model yet, will cause yet another revision to the model, and cause another big hit to the national accounts.

Second, at that point the government will continue to have three options – better information to applicants, caps on provision (or provider) type for the future, or further squeezing the students of the past by fiddling with repayment terms and screwing the students of the present by holding down maintenance, forcing them into full-time jobs while pretending to do full-time study and wrecking their skills and mental health in the process. Which in turn will impact their earnings and repayments, and so on.

Third, if this sort of stuff doesn’t come up at the Treasury Committee’s inquiry into student loans, I think I’m going to scream. Opacity over the Treasury’s calculations, considerations and motivations is doing untold harm, and call me old fashioned, but I think we have a basic right to influence these decisions through our politics.