Skip to content

Expected Shortfall Dashboard

Live risk monitor — instantiates Tail-Dominance Modeling and Control

Reports the average loss beyond a high quantile — not just the quantile itself — and tracks that tail average over time to catch the tail worsening.

A quantile like Value-at-Risk tells you a line the loss will rarely cross, but says nothing about how bad things get once it is crossed — two exposures with the identical VaR can have wildly different catastrophe potential. The Expected Shortfall Dashboard swaps that blind metric for one that averages the region beyond the threshold: expected shortfall (a.k.a. conditional VaR) is the mean loss on the worst, say, 2.5% of outcomes. Its defining move is that it reads the severity of the tail, not just its edge, and it does so as a living surface — trending the tail average so a worsening tail shows up on a screen before it shows up in a loss. It measures and watches; it does not model the tail's shape or set the limits (those are its siblings' jobs).

Example

A trading desk reports one-day 99% VaR to its risk committee: "we won't lose more than ≈$4M on 99% of days." The number has been flat for weeks, so everyone relaxes. But VaR is silent about the 1% of days it excludes. The desk moves its board to expected shortfall — the average loss across the worst 2.5% of days — the measure banking regulators themselves shifted to when the Fundamental Review of the Trading Book replaced 99% VaR with 97.5% ES.

Over three weeks the dashboard shows ES climbing even while VaR stays flat: a position has quietly concentrated, so the shape of the tail deepened without the 99% point moving. The drift alert fires. The desk investigates and trims the concentration before a bad day arrives — acting on a rising average-beyond-the-line rather than waiting for the line itself to break.

How it works

The metric is a conditional average: take all outcomes worse than a chosen quantile and average their loss, so the number responds to how heavy the tail is, not merely where it starts. That single value is then carried as a time series against a trailing baseline (or a budget line), and a rise beyond a set tolerance raises a drift alert. Showing ES and VaR side by side is deliberate: the moments they diverge — VaR flat, ES rising — are exactly the tail-deepening events a quantile alone hides.

Tuning parameters

  • Tail cut (quantile) — where "beyond" begins (95% vs 97.5% vs 99%). A deeper cut watches rarer, more severe events but rests on fewer observations and reads noisier.
  • Lookback window — how much history the ES estimate and drift baseline use. Short reacts fast to a regime change but jumps around; long is stable but slow to flag a worsening tail.
  • Drift-alert threshold — how large a rise over baseline trips an alert. Tight gives early warning at the cost of false alarms; loose is quiet but warns late.
  • Metric pairing — ES alone, or ES and VaR together; showing both surfaces the VaR-flat / ES-rising divergence that is the whole point.
  • Aggregation — one portfolio-total ES, or a per-book breakdown that localizes where the tail is deepening.

When it helps, and when it misleads

Its strength is that it reports a number sensitive to how bad the worst cases are, and it is coherent in a way VaR is not — combining two exposures never makes measured risk look larger than the parts, so it does not punish diversification.[1] As a live surface it turns tail-worsening into something you can see and act on early rather than reconstruct after a loss.

Its limits are the tail's own thinness of data: deep in the tail there are few observations, so the estimate is noisy and leans on modeling assumptions the dashboard displays but cannot certify. A number on a screen also breeds false comfort that the tail is "handled." The classic misuse is running the dials backwards — quietly moving the quantile or lengthening the window until the figure sits under a limit. The discipline that guards against it is to carry the estimate's uncertainty, stress it against a heavier-tailed model, and treat a rising ES as a prompt to act, not a parameter to re-tune.

How it implements the components

  • decision_metric_replacement — it retires the average- or quantile-based decision metric (mean, VaR) in favour of a tail-aware one (expected shortfall) that the appraisal actually steers on.
  • tail_drift_monitor — it trends that metric over time and alerts when the tail average worsens, making drift visible continuously rather than at post-mortem.

It does not fit the model that says where the tail begins (tail_boundary_definition, model_scope_boundary — Extreme-Value Threshold Model), set the loss budget it is read against (tail_consequence_budget — Heavy-Tail Simulation Scenario Set, enforced by Exposure Cap Policy), or diagnose the tail's shape (tail_shape_diagnostic — Tail-Index Estimation).

  • Instantiates: Tail-Dominance Modeling and Control — it is the standing metric surface that keeps the governed tail in view.
  • Consumes: Extreme-Value Threshold Model supplies the deep-tail estimates behind the metric when raw history runs out.
  • Sibling mechanisms: Exposure Cap Policy · Extreme-Value Threshold Model · Cumulative Contribution Curve · Heavy-Tail Simulation Scenario Set · Log-Log Survival Plot · Rare-Event or Importance Sampling · Reserve Buffer Policy · Robust Tail Statistic Review · Stress Test and Reverse Stress Test · Tail Incident Review · Tail-Index Estimation

Notes

Expected shortfall is only as trustworthy as the tail data behind it. At deep quantiles (99% and beyond) there are usually too few observations to average directly, so the dashboard's most extreme numbers rest on a fitted tail model rather than raw history — the surface displays the figure but does not vouch for it. Read the deepest cells as model output, not measurement.

References

[1] Expected Shortfall (Conditional VaR) is the expected loss conditional on being beyond a quantile. Unlike VaR it is a coherent risk measure — in particular subadditive — so a diversified book is never scored as riskier than the sum of its parts, which is one reason regulators moved trading-book capital from VaR onto ES. ↩