ESCRC Calculator

Frequently asked questions

Answers to the questions people ask most after their first run. Every figure below is a model-implied estimate for decision support — screening and ranking suppliers — not a capital-setting calculation.

Why doesn't the model see my biggest single-source supplier risk?

It depends on whether that supplier has a declared criticality (κ). Without one, the model prices a supplier's pro-rata revenue exposure — how much of your spend flows through it — rather than its criticality, how much of your output stops when it stops. A sole-source $7 component and a commodity of the same annual spend then produce the same tail, which understates the sole-source case badly.

The clearest historical case: the February 1997 fire at Aisin Seiki, which made 98% of Toyota's brake-fluid proportioning valves. The part cost about $7. Toyota stopped 20 of 30 lines and lost roughly ¥160bn of revenue, ¥20–30bn of it never recovered — against that supplier's ~0.04% share of Toyota's COGS, a spend-based model badly understates an event like this.

What to do about it: declare κ and β for the suppliers you care about — both fields are in the supplier table, and AI research will propose a κ with a source or refuse to guess one. Once declared, severity is driven by κ (the fraction of your output the supplier's stoppage halts) and β (the fraction never made up), not by spend share — a sole-source $7 component with a high κ then produces a far larger tail than a same-spend commodity with a low one. Until you set κ, treat the model's figure for a known sole-sourced, long-requalification supplier as a floor, not an estimate.

Tail Intensity shows only a few days for a supplier that was down for months historically. Is that wrong?

The number is right — it just isn't downtime. Tail Intensity is the supplier's worst-year loss at your chosen confidence level, divided by its daily revenue exposure, expressed in loss-equivalent days: the tail-year loss as equivalent days of the full revenue flow through that supplier. It's smaller than a calendar outage for two reasons: the impact coefficient scales a disrupted day by how much of it inventory buffers and second sources absorb, and it's an unconditional annual tail — the worst 1% of model years, not a "given the disaster happened" scenario. Use Tail Intensity to rank suppliers per dollar of dependence, never to read as predicted days of outage.

What does VaR mean in terms of revenue and profit?

Every VaR/ES figure in the tool is revenue at risk — top-line revenue foregone, not profit. Portfolio VaR at your chosen confidence level is the revenue loss the network stays below in most model years; the worst years beyond it average out to Expected Shortfall (ES). Standalone VaR is the same worst-year figure for one supplier alone — standalone figures are not additive, since independent suppliers rarely all have a bad year at once, so rank with standalone and total with portfolio. Profit at risk isn't output directly: a rough floor is VaR × gross margin, and the ceiling is the full revenue VaR when costs are fixed over the disruption window.

Why did raising the correlation setting (ρ) make VaR go down? Is that a bug?

No — it's a real property of correlation, not a defect. Regional correlation makes suppliers in the same geography more likely to have their bad and good years together, which pushes probability toward both ends of the loss distribution at once: more quiet years and more co-disaster years. A fixed-confidence VaR can fall if the extra quiet years outweigh the fatter tail at that specific percentile — the textbook reason VaR isn't a fully coherent risk measure. Expected Shortfall doesn't have this problem — it's mathematically guaranteed not to decrease as correlation rises. When comparing runs at different correlation settings, use ES, not VaR.

What's the difference between the parameter band, the confidence level, and the VaR confidence interval?

Three different numbers that happen to share a screen. Confidence level (e.g. 95%) picks which loss quantile counts as VaR — a risk-tolerance setting, not a statistical confidence statement. The VaR confidence interval measures Monte Carlo sampling noise — how much the reported VaR would move across random seeds at your sample size; it shrinks as the simulation runs more paths and says nothing about whether your inputs are right. The parameter band measures input uncertainty — how much VaR/ES would move if the underlying disruption-rate, recovery-time and impact estimates were different within their disclosed ranges; it does not shrink with more simulation paths, only with better evidence on the inputs themselves.

My VaR's lower and upper confidence-interval bounds are the same number, marked "pinned to atom." What does that mean?

Your network's quantile has landed on a discrete loss plateau, and the interval really is zero-width there — not the model claiming false precision. Loss in this model isn't continuous: a supplier that fails a whole number of times, for a fixed recovery time and impact coefficient, produces a specific loss value that can carry a large slice of probability on its own. When the quantile sits inside one of those, both interval edges land on the same value. It tells you the tail of that network is dominated by one supplier's discrete event steps rather than sampling noise — don't read the zero width as precision, and don't compare VaR across correlation settings on a network like this (use ES instead).

What still limits the model, even with criticality declared?

• Availability only. The model prices output a disruption gates. It can't express input-price or pass-through risk — a supplier that keeps shipping at triple the price is invisible to it.
• Bounded tail. The simulated loss distribution's extreme quantiles are structurally shorter than disclosed historical disruption losses — the standing reason a capital-setting claim is not made for this tool.
• Frequency is literature-anchored, not site-specific. Read the ordering of suppliers with more confidence than the absolute level of any single figure.