Exciting times! HM Treasury has published its Supplementary Estimates for 2025-26.
These are the documents through which government departments ask Parliament for permission to spend more money than was originally planned – topping up what was agreed earlier in the year when circumstances have changed.
Buried inside the Department for Education’s figures are some very big numbers that tell a very bleak story about the already beleaguered student loan system in England.
The short version is this – the government now expects to get back significantly less of the money it has lent to students than it previously assumed. The long-term cost of the system has been revised sharply upwards.
And the implications stretch from Whitehall accounting all the way to what comes out of every graduate’s pay packet – and potentially to big questions about whether the system can ever be made to work.
This is eye-wateringly complex stuff, so settle in.
How student loans work in public accounts
Before the numbers can make any sense, it’s worth understanding how student loans are treated in government bookkeeping.
When the government lends money to a student, it doesn’t treat the whole amount as “spending.”
Instead, it records three things separately.
- First, the cash going out the door – treated as a loan, not as spending.
- Second, an estimate of how much of that lending will never come back – the expected write-off. That’s the bit treated as a cost.
- Then third, the bit it does expect to get back. That shows up as an asset – something the government owns – on its balance sheet.
Think of it like this. If you lent a friend £100 and expected to get £60 back, you would record £100 as a loan, £40 as a loss, and a £60 asset.
The government does essentially the same thing, except the amounts are in the tens of billions and the “how much comes back” estimate depends on complex forecasts stretching decades into the future.
The problem is that those forecasts keep getting re-estimated as new information arrives about graduate earnings, repayment behaviour, economic conditions, and technical assumptions. When the government gets more pessimistic about how much will be repaid, the “cost” side gets bigger and the “asset” side gets smaller.
That re-estimation is what has just happened – at a very significant scale.
The big numbers
Earlier in the year, DfE’s budget line for these kinds of non-cash adjustments to the loan book was negative – at around minus £4 billion. In the Supplementary Estimates, it swings to a positive £5.7 billion. That is a movement of roughly £9.7 billion in-year.
The single biggest driver within that swing is a line worth just over £4 billion, described in Treasury jargon as being about “effective interest, discounting, and other changes.” Strip that away, and it’s the government admitting that the student loans it holds are worth less than it previously thought.
Separately, there’s a £19.2 billion adjustment to what’s called the RAB charge – the government’s running estimate of what proportion of every pound it lends to students will never come back. That estimate has gone up substantially. If the loans you hold are worth less than you thought, you have to record the difference as a cost – and that’s what this line does.
And the system continues to issue new loans at scale. DfE’s budget for new lending is over £30 billion. The loan book is still growing, even as the expected recovery rate is falling.
In other words, we are adding new lending that is expected, on average, to return less to the taxpayer than previously assumed.
None of this in and of itself changes the cash students receive today, the fees universities charge, or the repayment rules graduates currently face. It is about recognising in the public accounts that the system costs more than was previously assumed.
And the impact shows up as higher public spending in the year the assumptions change, rather than being spread smoothly over decades.
Whodunnit
Let’s go a little further down the rabbit hole. The distinction between the two main candidate explanations matters a lot – because they have very different implications for what happens next.
Explanation one – the discount rate changed
When the government estimates what a stream of student loan repayments stretching decades into the future is “worth” today, it doesn’t just add up the expected cash. It applies something called a “discount rate” to work out what those future payments are worth in today’s money.
The concept is straightforward. A pound you receive in 30 years is worth less than a pound you receive today, because you could invest today’s pound and earn a return in the meantime. The discount rate captures that difference. If the rate goes up, future money is treated as less valuable. If it goes down, future money is treated as more valuable.
For student loans, this matters enormously – because repayments can stretch over 30 or 40 years. Even tiny changes in the discount rate compound over that time horizon and can move the measured value of the loan book by billions, even if nothing has changed about what graduates actually earn or repay.
And the rate has changed. Between March 2025 and March 2026, the Treasury increased the discount rate used for student loans by 0.30 percentage points. That sounds tiny – but when you’re discounting repayments stretching decades into the future, small changes compound into massive differences.
To put that in perspective – the National Audit Office reported that at 31 March 2025, the loan book had a face value of £265 billion and was valued in the accounts at £158 billion. Roughly speaking, a 0.30 percentage point increase applied across a loan book of that size could knock something like £7 to £10 billion off its measured value.
That is “mechanically possible” scale for a discount-rate-driven movement, and it is in the same universe as the multi-billion re-estimate adjustments we are seeing.
The language in the Supplementary Estimates supports the reading. The £4 billion line explicitly bundles “effective interest” and “discounting” as drivers. That is valuation-mechanics language – it points to the discount rate machinery rather than to a sudden collapse in observed repayments.
I should be clear about what that means. The discount rate isn’t the government’s actual borrowing cost – it’s not the interest rate paid on government bonds.
It is a valuation tool used to convert future money into today’s money. It would exist even if the government were running a surplus and not borrowing at all. But it is influenced by prevailing interest-rate conditions, which is why it moves. And when it moves, the effect on a decades-long cashflow stream like student loans can be dramatic.
Explanation two: graduates are not repaying as much as expected
The second candidate explanation is also a problem, because it implies something is structurally wrong with the assumptions underpinning the system rather than just a technical parameter shift.
As I noted a while ago, 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.
Audit Scotland’s annual audit report for the Scottish Government said the model’s methodologies:
…are believed to over-forecast student loan repayments, leading to understated impairment charges.
Both auditors were issuing formal warnings – and in the Welsh case, an actual audit qualification – that the existing model was seriously overstating what governments would get back in repayments.
Both noted that DfE was building a new repayment model, expected to use better data and methods. Audit Wales explicitly said adopting the new model was “likely to result in a substantially lower student loan asset value in 2025-26.” And here we are.
How much is each?
We don’t know. The Supplementary Estimates don’t break out the £4 billion re-estimate into “discount rate changed” versus “repayments worse than modelled.” But in the absence of transparency, it’s worth having a think about why the model might have been wrong.
The DfE student loans forecasting system is not a single model. It is a chain of linked models.
There’s an entrants model (how many people borrow), an outlay model (how much they borrow), an earnings model (what they go on to earn), and a repayments model (what they actually pay back).
Together they generate the forecasts used for budgeting, policy costings, and valuing the loan book in DfE’s annual accounts.
At its core, the system builds individual borrower profiles, predicts an earnings path for each borrower year by year, then applies the repayment rules – thresholds, rates, interest, write-off terms – to generate projected cashflows.
It is technically sophisticated. But it has a structural problem.
The two-stage earnings model
For the first ten years after a graduate starts being required to pay money back if they hit the threshold, DfE uses an “early-career” earnings model.
It’s trained on Student Loans Company data and a linked dataset called LEO that tracks graduates’ actual tax records. At this stage, things like subject of study shape the projected earnings. Graduates from different subject groups start in different earnings brackets, and that is reflected.
But provider characteristics – where you studied – enter only weakly, through broad groupings like “high tariff, medium tariff, low tariff” rather than institution-specific assumptions. There is no “this university produces this earnings trajectory” in the model.
Then for years 11 to 43, DfE switches to a completely different model. This one is trained on a sample of general HMRC earnings data – the whole population, not just graduates.
The assumption is that by your mid-thirties, what you studied and where you studied it no longer meaningfully determines your earnings. Your actual earnings history takes over as the predictor.
This means that subject and provider largely drop out of the model after the first decade – whatever you did at university is assumed to have faded as a factor by then.
What happens when the mix of students changes
The problem is that the model doesn’t adjust its assumptions when the mix of who enters higher education, and the mix of what they study, shifts.
If participation expands disproportionately among courses and providers that generate persistently weak earnings outcomes, the model doesn’t automatically downgrade its long-run expectations for those borrowers. It just keeps projecting earnings based on what earlier, different cohorts achieved.
Let me put it like this. If a university historically produced graduates who earned well, and it then expanded rapidly through franchise partnerships where outcomes were systematically weaker, the model would just assume the new graduates look like the old ones. It matched new borrowers to the observed trajectory of similar-looking borrowers – but if the category was itself being diluted by expansion, the match becomes misleading.
Or let me put this another way. If lots of students move up the tariff ladder – from low to medium, and from medium to high – the model thinks this is good news because more students are treated as future high earners.
But medium doesn’t really grow. It loses students to high tariff and replaces them with students coming up from low tariff. Both medium and high tariff then may end up containing more students who earn less than the old students in those groups did. The model still assumes everyone in each group earns like the old groups. So it expects too much money back. In reality, repayments are lower and come later, and the system ends up costing far more than the model predicted. The model being wrong, on the changes we’re seeing, has been baked in.
DfE’s own 2024-25 accounts acknowledge a related problem. They note that their own models don’t agree with each other on when repayments arrive – which matters, because under discounting, when money comes back is almost as important as whether it comes back at all.
The model can therefore be simultaneously “right” about next year’s total repayments – because those are dominated by established graduates where the broad economic assumptions do most of the work – and structurally wrong about lifetime repayments, which depend on the long-run earnings trajectories of newer, different cohorts.
That is a model that looked fine on short-term accuracy checks while quietly accumulating a hidden liability. It could overestimate how many borrowers ever repay meaningfully, overestimate late-career repayments, and therefore overstate the recoverable value of the loan book.
That’s the failure the devolved auditors appeared to be identifying when they said repayments were running at half the forecast level.
Three sets of books, one problem
Part of the reason this story is so hard to follow is that three different accounting systems are trying to describe the same underlying reality – and they use different methods, different assumptions, and produce different numbers.
System one: DfE and HM Treasury (departmental budgets and accounts)
This is where the Supplementary Estimates sit. When each pound is lent, DfE splits it into a “recoverable asset” and a “cost” using a discount rate set by the Treasury. Changes in that discount rate – even without any change in what graduates actually repay – can move the measured cost by billions.
This system asks:
If we lend this money to students instead of using it for something else, what does that cost us?
System two: the Office for National Statistics (national accounts)
ONS uses a different method. When loans are issued, it decides how much of the lending is a genuine financial asset (expected to be repaid) and how much is really a grant dressed up as a loan (expected never to be repaid).
Crucially, ONS bases that split on the interest rate charged to borrowers, not the government’s own cost of funds. So a Treasury discount rate change doesn’t automatically change the ONS numbers.
This system asks:
Is this transaction economically a loan or a transfer?” – not “What does it cost the state?
System three: the fiscal rules
The Chancellor’s fiscal rules are assessed on ONS figures, not on DfE budget lines. So the very large movements in the Supplementary Estimates don’t automatically translate pound-for-pound into a hit to headline fiscal targets.
But if they reflect a genuine deterioration in expected repayments – rather than purely technical discounting effects – the ONS numbers will eventually catch up, because ONS revises its figures when expectations change.
The result is that the same policy can look cheaper or more expensive depending on which set of books you’re looking at.
Discount rate changes can cause accounting earthquakes in DfE’s budgets without immediately showing up in headline fiscal numbers. Repayment underperformance can take years to feed through from observed data into model revisions into accounting adjustments into fiscal figures.
And at every stage, the numbers are large enough to matter – but opaque enough that nobody outside a small circle of specialists notices.
This is not an accident. Each system is protecting a different objective. DfE’s system promotes cost awareness. The ONS system ensures statistical comparability. The fiscal rules try to sit across both and end up inheriting the blind spots of each. But it pretty much prevents any meaningful public scrutiny of what’s going on.
What it means for the Chancellor’s fiscal rules
Rachel Reeves’s fiscal framework includes an “investment rule” that requires something called public sector net financial liabilities (PSNFL) to be falling as a share of GDP by the target year of the forecast – currently 2029-30.
We used to just measure the amount of borrowing we were doing. Now it’s a bit more sophisticated – PSNFL is a measure of the government’s overall financial position – everything it owes minus everything it owns.
Student loans are one of the biggest financial assets in that calculation. If the expected value of the loan book falls, the asset shrinks, net liabilities rise, and PSNFL gets worse.
Reeves likely chose this measure partly because it makes student loans look good on the books – the portion expected to be repaid counts as a government asset, making the public finances look healthier than under the previous debt metric.
But it also means the rule is sensitive to revaluations of those assets – and student loans are the single biggest and most volatile asset in the framework.
The Institute for Fiscal Studies has warned about this. A rule based on PSNFL can create weird incentives – favouring policies that create financial assets (like student loans) over alternatives (like grants or a graduate tax) purely because of how they affect the balance sheet, rather than because they are actually better policy. And it is vulnerable to sudden shocks when the value of those assets gets revised.
At the November 2025 Budget, the OBR reckoned that the government was on course to meet the PSNFL rule with headroom of about £24 billion – roughly 0.7 per cent of GDP – and gave it a 52 per cent probability of success. That is barely better than a coin flip.
If the £4 billion re-estimate in these Supplementary Estimates were fully reflected in the loan book’s value, PSNFL headroom would fall from roughly £24 billion to roughly £20 billion. The rule wouldn’t be breached – but the margin would be tighter, with the probability of meeting it falling further below even odds.
That is a rough estimate, not a precise number – the two accounting systems don’t map onto each other pound-for-pound. But the direction is clear – a weaker student loan asset means worse PSNFL, and worse PSNFL means less headroom against the Chancellor’s own fiscal rule.
And it would have been worse without that repayment/interest rate freeze that was buried in the Budget that the broadsheets have belatedly noticed.
The policy incentive this creates
If you are Chancellor and your fiscal rule depends partly on the value of the student loan asset, and that asset is being written down, you have a very direct incentive to increase the expected value of repayments. That is how you prop up the asset and protect your headroom.
And the simplest lever available is freezing the thresholds.
When the repayment threshold is frozen instead of rising with earnings or prices, more borrowers are pulled above the repayment line. A larger proportion of their earnings becomes subject to repayment. Repayments arrive earlier in the life of the loan – and earlier cashflows are worth more under any discounting approach. The expected value of the loan asset rises. PSNFL improves. Similar things happen when you freeze the stepped interest rate.
As you’ll have noticed, the 2025 Autumn Budget included a freeze in the Plan 2 repayment threshold, and a freeze in the associated stepped interest rates – and the IFS said the main fiscal effect was to make the existing pile of student loans look more valuable on the government’s balance sheet.
Plan 5 – the repayment plan for students entering from 2023-24 – can be read through the same lens. Its design features a lower repayment threshold than Plan 2, a forty-year repayment term, and threshold indexation terms that are less generous than earnings growth. All of these mechanically reduce the proportion of lending the government expects to lose.
If the Treasury anticipated that discount rate changes were about to worsen loan-book valuations, tightening terms in advance makes sense as fiscal management – regardless of whether it is fair to graduates.
There is no public evidence that the planned freezes were explicitly designed as responses to anticipated valuation problems. But the alignment between “things that improve PSNFL” and “things the government has actually done to repayment terms” is close enough to be noteworthy.
The fairness problem
The problem is the logic of what is happening.
The government expands higher education, but shifts a good wedge of the cost of doing so onto graduates – lending large sums to students on the promise that a degree will improve their earnings via a “graduate premium”, but basing that promise on what happened in the past.
Then the graduate premium starts to fall. Part of that is the floor – non-graduates’ earnings rise, partly via the national minimum wage. That erodes the original case for asking graduates to shoulder the share they were asked to – £9k fees, etc.
The repayment model assumes graduates will earn enough to repay a given share. But then the government notices that its model is wrong, either:
- Because the economy in general underperforms, or
- Because of “dilution” of the benefits in general (more graduates), or
- Because the system expands provision in places or courses where graduates end up earning less (franchised provision in business courses in office blocks), or
- Because universities get worse at providing the education and skills the economy needs, or
- A mix of all four.
Over time, the model’s assumptions about repayments turn out to be too optimistic. The loan book loses value. Fiscal headroom shrinks. And the government responds by tightening repayment terms – freezing thresholds, extending repayment periods – so that graduates pay more from lower incomes.
That means, if you think about it, it’s even worse than Martin Lewis says. Because graduates end up being sold a three-pronged pup.
- First, they were asked to pay more than in the past to fund expansion – so the initial terms are worse than in the past.
- Then, their own earnings aren’t as good as imagined, which costs them more than they imagined.
- Then, the Treasury fiddles with the terms to make them pay even more.
Put another way – a system which was supposed to protect graduates from poor labour market outcomes fails to do so, but they then have to pay for the failure. Twice.
The bigger problem, though, is this.
It is now increasingly difficult to simultaneously meet HMT’s expectations of how much money comes back, to meet students’ and providers’ expectations of how much money they need (at least on current student numbers), and maintain a progressive repayment system that is also politically acceptable to graduates. It is not a circle that can be squared.
It’s like those old Penelope Pitstop cartoons where all the walls are closing in. As the graduate premium tightens – whether because of better non-graduate earnings, a weak economy, dilution through too many graduates, the wrong subjects in the wrong bits of the country, or poor teaching leading to poor skills – the justification for asking graduates to pay weakens.
But from HMT’s perspective, the tighter the premium gets, the more it needs to extract from graduates at lower wages, or the fiscal numbers do not add up. The argument for raising fees gets weaker at exactly the moment the pressure to collect more gets stronger.
What we still don’t know
As I say, the Supplementary Estimates don’t split up the £4 billion re-estimate into “discount rate changed” versus “repayments worse than modelled.”
DfE’s own analysis suggests that earnings assumptions matter more than discount rate assumptions in driving the measured cost of the system. But the Estimates lump everything together without telling us the split.
We don’t therefore know the distribution of the repayment shortfall that auditors in Wales and Scotland have flagged. Which cohorts are underperforming? Which providers? Which subjects? Which nations?
If the shortfall is driven disproportionately by certain kinds of expansion, it becomes important to control that expansion – or at least use some kind of planning tools.
Put another way, you might be opposed to student number controls for good reasons. But without them, at the moment we are literally making graduates pay for the failures of uncontrolled marketisation.
I might have a bit of the above wrong, or implied one factor is more important than another. But that speaks to a wider issue – someone in government needs to explain publicly what is happening, who is affected, and what it means for the deal between graduates, universities, and the state.
Because right now, the numbers are moving in the accounts while the politics remains silent. And graduates are picking up the tab.