FRTB: The Fundamental Review of the Trading Book

During the 2008 crisis, banks' trading-book losses ran far past what their VaR models had signaled was possible. Part of the problem was the metric itself: a 99% VaR tells you almost nothing about how bad the losses get beyond that 99th percentile, and it was exactly that tail that banks got hit by. FRTB, finalized by the Basel Committee, is the rulebook written to fix that measurement gap, and to close several other loopholes the pre-crisis framework left open.

The headline change: VaR out, expected shortfall in

Under the old Basel framework, trading-book capital was based on Value at Risk (VaR) at 99% over a short horizon: a single number saying "we don't expect to lose more than this on 99 days out of 100." The problem is that VaR is silent about the worst 1%. Two banks can report identical VaR figures while one has a modest tail beyond it and the other has a catastrophic one; VaR can't tell them apart.

FRTB replaces this with expected shortfall (ES) at a 97.5% confidence level: the average loss in the worst-case scenarios beyond that threshold, not just the boundary. See Expected Shortfall (CVaR) for the mechanics. In words, instead of asking "what's the loss level I'm 99% confident I won't breach," FRTB asks "given that I'm in the bad 2.5% of outcomes, how bad does it average out to be." That second question can't be dodged by a portfolio whose losses are fine most of the way to the boundary and then explode past it.

Worked example

A trading book has a distribution of one-day P&L outcomes. Sorted from worst to best, the bottom 2.5% of scenarios (in a 400-scenario historical simulation, the worst 10 days) show losses of: $12m, $11m, $10m, $9.5m, $9m, $8.5m, $8m, $7.5m, $7m, $6.5m.

• 97.5% VaR looks only at the boundary: the 10th-worst loss, $6.5m. It says nothing about how much worse the other nine scenarios were.
• Expected shortfall averages the whole tail: $(12+11+10+9.5+9+8.5+8+7.5+7+6.5)/10 =$ $8.9m.

The ES figure is 37% higher than the VaR figure precisely because it's picking up the fatter, worse losses that VaR ignores past the boundary. A desk with a long tail of rare, severe losses will see its capital charge jump under FRTB even if its VaR was identical to a desk with a thinner tail — which is the entire point of the change.

Two other structural changes

A stricter trading book/banking book boundary. Before FRTB, a bank had some latitude to place an instrument in whichever book required less capital, the trading book for liquid assessments, the banking book (held-to-maturity) for less liquid ones with more favorable treatment. FRTB tightens the criteria and imposes a capital penalty for moving instruments between books after the fact, closing an arbitrage that let banks reclassify their way to lower charges.

Liquidity horizons instead of one flat window. Instead of assuming every position can be hedged or exited in ten days, FRTB assigns different liquidity horizons — 10, 20, 60, 120 days — by risk factor, so an equity index position and an illiquid credit tranche are no longer measured on the same unrealistic timeline. A position with a longer liquidity horizon carries a materially higher capital charge, reflecting that it would take longer to safely exit if the desk needed to.

What this means in practice

• Desks that previously looked cheap to run can become expensive under FRTB purely because their tail or liquidity profile was worse than their VaR let on — this has driven real changes in which products banks are willing to warehouse.
• Model approval got harder. Each trading desk's internal model must pass backtesting (see VaR Backtesting and the Kupiec Test) and a "profit and loss attribution test" desk by desk; a desk that fails falls back to a much more conservative, less risk-sensitive standardized approach.
• Expected shortfall is harder to backtest than VaR because, unlike a VaR breach, an ES estimate isn't a single testable exceedance — regulators and banks are still refining how to validate it robustly.