| Vose Software

Industry: Legal
Product: ModelRisk
Application: Contract Risk Assessment


A $60M reserve that looks safe is breached 11% of the time

A technology company's legal team holds a portfolio of 40 high-value commercial contracts and a $60M reserve set aside against breach, penalty, and dispute losses. Run the contracts independently — each with its own small chance of going wrong — and the maths is comforting: most years nothing much happens, the aggregate loss hovers near its mean, and $60M looks like generous cover. The deterministic spreadsheet that sums the expected loss of each contract agrees, returning an annual exposure around $24M.

That comfort is an artefact of an assumption no general counsel should make: that contracts fail independently. They do not. A recession, a stressed counterparty sector, or a supply-chain shock pushes many counterparties toward default in the same year. Once that shared driver is in the model, the aggregate loss develops a fat right tail the independent sum cannot produce. The simulation puts the 99% VaR at $135.7M — more than double the reserve — and the probability that annual losses exceed the $60M reserve at 10.7%, better than one year in ten. The point estimate did not lie about the average. It was silent about the tail, and the tail is the entire reason a reserve exists.

Building the portfolio loss, one contract at a time

Each contract is modelled as a frequency-severity pair, with parameters that vary across the book:

  • Breach trigger — a Bernoulli event. Baseline breach probabilities are low, drawn per contract in the 2%–9% range, reflecting that most contracts perform most years.
  • Loss given breach — a LogNormal severity in dollars, with per-contract medians spanning $1.5M–$12M and log-sigmas of 0.45–0.85. Legal losses are right-skewed: most breaches are manageable, a few are ruinous, and a Normal would understate the worst cases.

The portfolio loss in any simulated year is the sum over contracts of trigger x severity. If that were the whole model, the central limit theorem would quietly flatten the aggregate into a thin near-Normal hump — 40 small independent risks averaging each other out — and badly understate the reserve a regulator or auditor would demand.

Why independence is the dangerous assumption

The fix is a shared market-stress factor: one standard-normal draw per simulated year, common to every contract, representing the health of the economy and the counterparty sector. Each contract loads on it with its own sensitivity (drawn 0.6–1.4), and the loading is applied in log-odds space and squeezed back through a logistic so every breach probability stays a valid number in [0, 1]. In a stressed year the factor lifts every contract's breach probability together — so breaches cluster, and a bad year is one where many contracts fail at once rather than just one.

This is the move that creates the tail, and the model verifies it empirically rather than asserting it:

  • Mean pairwise correlation of the breach indicators under the shared-stress model: 0.081, versus 0.000 under independence — the factor demonstrably induces positive co-movement.
  • The per-contract marginal breach rate is 0.082 in both models — the shared factor reshapes the dependence without disturbing any single contract's own odds, so the two models are a fair like-for-like comparison.
  • The aggregate 99% VaR rises from $86M (independent) to $135.7M (correlated) — correlation fattens the tail by 58%.

Aggregate portfolio loss distribution with VaR and reserve

The headline distribution makes the asymmetry plain. Most years cluster near the $24.4M mean, but the tail runs far to the right: VaR 95% = $84.2M, VaR 99% = $135.7M, and the expected shortfall beyond the 99% point — the average of the worst 1% of years — is $167.0M. The dotted line marks the $60M reserve, and the simulation reads off the answer that matters directly: P(loss > $60M reserve) = 10.7%.

The reserve question, drawn as an exceedance curve

The cleanest way to size a reserve is to plot the probability of exceeding every possible loss level — and to overlay what the same portfolio would look like if breaches really were independent:

Loss-exceedance curve, correlated versus independent

The two curves share the same body but diverge dramatically in the tail. At the $60M reserve the correlated model exceeds it 10.7% of the time. Push out to a $100M loss and the gap is stark: the shared-stress portfolio breaches $100M 3.0% of years, the independent portfolio only 0.4% — correlation multiplies the tail risk by roughly 7x. A reserve sized on the independent (dashed) curve would be set against a one-in-250-year event that is really a one-in-33-year event. The VaR markers sit where they should: 95% at $84M, 99% at $136M, both far beyond the reserve a naive sum would have justified.

Three credit cycles, three loss curves

Shifting the mean of the shared stress factor sweeps the portfolio through a calm, baseline, and stressed credit environment — a genuine re-simulation at each setting, not a rescaling:

Three credit-environment scenarios

  • Calm credit cycle — mean $14.3M, P95 $55M, P99 $95M
  • Baseline — mean $24.4M, P95 $84M, P99 $136M
  • Stressed credit cycle — mean $46.0M, P95 $135M, P99 $192M

In a stressed cycle the P95 alone ($135M) exceeds the baseline's P99 — the same 40 contracts, the same severities, simply more synchronized. This is the scenario a provisioning committee needs to see: the reserve that covers 95% of baseline years covers barely 90% of stressed ones.

What drives the portfolio loss

Ranking each contract's loss — plus the shared stress factor — by its rank-correlation contribution to the aggregate spread answers where the risk concentrates:

Tornado of portfolio-loss drivers

  • Shared market stress factor — ±$21.8M, the single largest driver by a wide margin
  • The six largest individual contracts — each contributing ±$8M–$11M, led by Contract 10 (median $7.2M severity) at ±$11.4M

The most important "input" is not any one contract but the common stress factor — the very dependence the independent model threw away. No amount of renegotiating a single agreement addresses it; only portfolio-level measures (diversifying counterparties across sectors, raising the reserve, or buying a portfolio-level guarantee) move the number that matters.

What the model changed

The legal team arrived with a $60M reserve and a $24M expected-loss spreadsheet that made it look ample. The simulation showed the reserve is breached 10.7% of years, that true 99% VaR is $135.7M with a $167.0M expected shortfall beyond it, and — most importantly — that 58% of that tail comes from correlation the independent model could not see, with the shared stress factor the dominant driver at ±$21.8M. The team could now size the reserve against the 99% VaR rather than the mean, and could justify portfolio-level diversification as the only lever that touches the top risk driver.

ModelRisk functionality used

  • Frequency-severity contract model — per-contract Bernoulli breach triggers (2%–9% baseline) combined with LogNormal severities (medians $1.5M–$12M) to build the aggregate portfolio loss.
  • Shared-factor correlation — a single annual stress draw applied in log-odds space across all 40 contracts, lifting the mean pairwise breach correlation from 0.000 to 0.081 while holding each contract's marginal breach rate fixed at 0.082.
  • Tail risk metrics — 95% and 99% VaR ($84.2M / $135.7M), expected shortfall ($167.0M), and the direct reserve-exceedance probability (10.7% above $60M).
  • Exceedance-curve comparison — plotting correlated against independent survival functions to quantify the 7x tail multiplier at the $100M level.
  • Scenario simulation — re-running the portfolio under calm, baseline, and stressed credit cycles to show the P95 of a stressed year ($135M) overtaking the P99 of the baseline.

Monte Carlo turns contract risk from "what is the average exposure?" into "how correlated are our counterparties, and how much reserve does that correlation actually demand?" — and the answer, here, is more than double what the average suggested.