The VAR-centric models that never were

Often spotlighted, rarely dominant – VAR plays a surprisingly small role in most IMA stacks

If prudential regulation were a political contest, value-at-risk would be the ultimate maverick – a 30-year survivor that’s repeatedly extended its term crisis after crisis, thanks to its adaptability and a fractured front among its challengers.

Since its introduction in 1996, VAR-based capitalisation of banks’ market risk has withstood multiple efforts to dethrone it. Not even the seismic collapse of 2007–08 managed to knock it out. Instead of scrapping VAR, the Basel Committee opted to complement it – adding a stressed VAR overlay and a handful of other ad hoc risk charges.

Those bolt-ons arguably allowed VAR to live to fight another day. But they also diluted its influence. Basel 2.5 reforms drastically reduced VAR’s share of internal model approach (IMA) capital requirements, shifting the burden to a set of less controllable, often bank-specific, gauges.

As a recent Risk Quantum analysis showed – drawing on quarterly IMA data from Q4 2007, covering 59 European and US banks – VAR-based risk-weighted assets (RWAs) accounted for just a median 19.9% of the total IMA stack at end-2024. Even at its peak, in Q2 2020, as trailing data inputs captured the worst of the Covid-19 panic, VAR never exceeded 31.6%.

In other words, for the median bank, VAR rarely accounted for even a third of modelled market risk RWAs – and usually much less. Most of the time, the bulk of the IMA output came from the complementary metrics introduced by Basel 2.5.

 

Stressed VAR (SVAR) leads the charge here, seldom falling below 50% of total IMA RWAs in any given quarter across the sample. Depending on the dealer, the incremental risk charge also encumbered sizeable shares of capital – a median 20.6% as of end-2024, with Deutsche Bank and Banco BPM coming in at 40.2% and 61.4%, respectively.

More striking still is the preponderance of RWAs that cannot be classified under any of Basel 2.5’s four building blocks. Generally, these are residual charges for model-blind spots – variously dubbed risks-not-in-VAR (RNIV), risks not in the model engines or specific risk add-ons, US rules, depending on the jurisdiction.

The bank with the heaviest such burden in our analysis was Oklahoma-based BOK Financial, with 67.1% of its IMA output falling into this residual category. But just behind were some of the heavyweights: JP Morgan at 61.9%, Bank of America at 60.5%, Citi at 57.7% and Morgan Stanley at 55.4%. That means that for some of the biggest global dealers, more than half of ostensibly modelled market risk capital stems from risks that escape modelling capabilities.

 

Portfolio complexity partly explains this – the broader the asset range, the harder it is to model every corner. But that doesn’t hold everywhere. ING Bank, UniCredit and Deutsche Bank – all dealers with diversified books – carried RNIV shares of just 3.9%, 3.1% and 2.7%, respectively. 

To be fair, the Basel Committee always viewed the Basel 2.5-era IMA as a mere stopgap – a transitional fix until its Fundamental Review of the Trading Book was ready to usher in a new era, built on expected shortfall. This measure shifts the focus from losses on 99% of trading days, as quantified by VAR, to the average loss in the worst 2.5% – in other words, tail risk.

But VAR still retains a key role under FRTB’s IMA. The multiplier that translates expected shortfall into capital requirements is calibrated via 99% VAR backtesting – just as under the old regime. At the same time, FRTB doubles down on the idiosyncratic and hard-to-model, formalising the RNIV framework – which had largely been the purview of local regulators under Basel 2.5 – through non-modellable risk factors.

For all the talk of VAR-centric models, the past 12 years of market risk capitalisation have never really been about VAR alone. For most IMA users, charges have been built on a hotchpotch of gauges and bolt-on adjustments.

Under the FRTB, that patchwork will persist – albeit with fewer banks qualifying to use internal models at all. But for those that do, the IMA will still look a lot like the one it’s replacing: a bespoke bundle of capital add-ons, where VAR retains a seat at the table – just not at the head.