The switch from 99% Value-at-Risk to 97.5% Expected Shortfall is the most consequential change to trading-book risk measurement in two decades. It changes not just a number, but the shape of the models - and of the data underneath them.
Why the regulator moved from VaR to Expected Shortfall
Basel 2.5 measured trading-book market risk with Value-at-Risk at the 99th percentile over a 10-day horizon. The measure had a known blind spot: it says nothing about how bad losses are beyond the quantile. A book can pass a 99% VaR test while carrying catastrophic exposure in the far tail. FRTB's answer was to switch to Expected Shortfall at the 97.5th percentile - the average loss in the worst 2.5% of outcomes - calibrated to a 250-day stressed calibration window rather than the most recent year.
The design choices are deliberate. The lower confidence level compensates for ES being estimated with less precision deep in the tail; the stressed calibration anchors capital to crisis-period behaviour, echoing the stressed-VaR concept of Basel 2.5 but making it the primary metric rather than a multiplier.
The machinery underneath: liquidity horizons and shifting
FRTB's most under-appreciated feature is its treatment of time. Each risk factor is assigned a liquidity horizon (LH) - from 10 days for liquid rates and major FX pairs out to 120 days for the most illiquid exposures - and the base-horizon ES is scaled to those horizons through the shifting-and-scaling approach. The practical consequences are severe:
- Concentration is now capital-relevant. Two portfolios with identical VaR can differ materially in ES once illiquid positions are scaled to longer horizons.
- Diversification is partially recognised, partially penalised. Cross-asset correlation benefits survive within a liquidity bucket but not across them - the exact point the March 2026 US re-proposal seeks to soften.
- Curvature risk moved to a separate charge (the sensitivities-based method), with vega and curvatures treated distinctly from linear delta risk.
Non-modellable risk factors: where models go to die
A risk factor passes the risk-factor eligibility test only if there are at least 24 (previously 10) real price observations per year. Everything else is an NMRF, and must be capitalised through a deterministic stress scenario (SES) built around a liquidity-horizon-scaled stress loss. In commodity and insurance books - exotics, weather derivatives, thinly traded credits - NMRF populations are routinely large, and the SES capital is often a multiple of what an internal model would produce. Reducing NMRFs (better data sourcing, proxying, and mapping) is usually the single highest-ROI capital project on an FRTB programme.
Backtesting ES: the academic frontier
ES is not elicitable - there is no natural scoring function to hit it with - which made the P&L attribution (PLA) tests and the spectral-weighted backtest that FRTB ships with feel, to many quants, like a workaround. The literature has moved fast: Acerbi and Szekely's ridge backtest offered a practical validation metric; Bayer and Dimitriadis' regression-based ES backtests (Journal of Financial Econometrics, 2022) gave the approach joint-testing rigour; and new exceedance-based methods published as recently as February 2026 continue to sharpen the toolkit. Spectral risk measures - families of coherent measures that weight the tail explicitly - are increasingly used internally to bridge between ES and the firm's own risk appetite.
What this means in practice
Three moves separate the firms that are ready from the firms that are re-acting: (1) build ES on a single governed simulation stack - one set of paths feeding VaR, ES, stress and PAA, not four reconciling models; (2) treat NMRF reduction as a data problem with a capital price tag; and (3) adopt modern ES backtests now, because supervisors will ask for evidence your ES model behaves long before they ask you to re-derive it.