A typical model has two masters, the deal team whose currency is comfort, but the credit committee deals in real risk. The model always satisfies the first master with DSCR, LLCR and PLCR, and ignores the second’s two questions: EDF and LGD. If these are new terms to you read on as they are fundamental to project finance risk and structuring.
Two deals land on your desk. Same DSCR profile, same LLCR, same PLCR, same tenor. You price them at similar margins and sign both. One repays in full. The other defaults in year three and you recover sixty cents in the dollar. Your model never saw it coming because the focus is on headroom not actual risk.
Coverage is comfort – pricing is risk – there is a disconnect.
DSCR, LLCR and PLCR answer one question asked three ways: how much headroom does CFADS have over debt service, this period, over the loan life, over the project life. I describe them as cushion metrics. They factor in the central case and the buffer around it – and for a range of scenarios. They do not surface the parameters that actually define credit risk: “how likely is a default”, and “how much do you lose when one happens”.
Rewind – remember that Expected Loss = Probability x Consequence. In credit risk terminology Expected Default Frequency (EDF) is the probability. Loss Given Default (LGD) is the consequence if that happens. Leaving them out means you are flying blind – not a good place to be.
The equation every credit team knows and no deal model contains
There is a bit of maths but it’s straightforward.
Expected Loss = Probability of Default(PD) * Loss Given Default (LGD) * Exposure At Default(EAD).
I think this is the #1 equation in credit risk, and find it’s not modelled in the hundreds of deal models I have seen, I do often as an interesting test but this does make me wonder if I am wrong and “it’s all fine”….
For an exposure over a defined horizon, expected loss is the product of three terms:
(Equation: expected loss equals PD times LGD times EAD, shown in the original post.)
Where:
EL = Expected Loss
PD = Probability of Default
LGD = Loss Given Default
EAD = Exposure At Default
The default probability is the EDF, recovered in the structural model from the distance to default, while loss given default is one minus the recovery rate:
(Equation: PD equals EDF equals Phi of minus distance to default, shown in the original post.)
Where:
PD = Probability of Default (as above)
EDF = Expected Default Frequency
Phi = Standard Normal Cumulative Distribution Function
DD = Distance to Default (How many Standard Deviations of headroom before the asset value hits the default point)
(Equation: the distance to default formula, shown in the original post.)
Where:
V = Asset Value (CFADS remaining at Default)
k = Where it defaults
Sigma = Asset Volatility (unlike listed assets, this needs some derivation but that’s’ doable)
T = Evaluation Horizon
(Equation: LGD equals one minus recovery rate, shown in the original post.)
Where:
LGD = Loss Given Default (as above)
R = Recovery Rate
And because project finance defaults are empirically front-loaded rather than uniform, the figure that actually prices the loan is the discounted sum over the tenor:
(Equation: lifetime expected loss as a discounted sum, shown in the original post.)
Where:
EL = Expected Loss (as above)
PD = Probability of Default (as above)
LGD = Loss Given Default (as above)
EAD = Exposure at Default (as above)
v = per period discount factor (cost of debt + for debate the default margin)
T = Tenor
t = period index (1 to T)
Open a typical model and you will find debt sculpted to a target DSCR, a covenant headroom test, a base case and a well-considered downside but not an expected loss number. The model that holds the entire CFADS distribution, the security package and the full tenor in memory declines does not the calculate the single figure that should price the loan. Always makes me scratch my head – I calculate it often just for fun to see how far off the pricing is.
If we were having a chat in the pub about this I could mount a fair objection to this hard to argue case: you cannot run a textbook EDF on a project. No listed equity means no market-implied asset volatility, and no distance to default to convert into a probability. True, but I’d say beside the point. The PD you need is already in your model, it is the frequency with which your simulated CFADS breaches the default point, and where there is no simulation the agency project finance default tables give a defensible term structure. The bank computes expected loss anyway, downstream in its capital model; my argument is purely that can also live upstream, in the cash flow model where the risk is actually made. If you now me you know I am a hawk for spurious precision, and I don’t think this is a candidate when managed pragmatically. Both PD and LGD on a long-dated bespoke asset are clearly uncertain, which is why lenders lean on legal and commercial structure and why the goal is a disciplined range, not a single hard number. DSCR already proxies the central case. What it cannot give you is the spread of outcomes around it, or what you keep when one goes wrong, and those are EDF and LGD – that’s all I am saying.
Project finance punishes this blind spot harder than any other sector
If there is one asset class where coverage and loss come apart, it is this one. The data, drawn from thousands of transactions over four decades, shows two things corporate credit doesn’t. Default hazard is front-loaded into construction and ramp-up, then collapses as the asset seasons and the contracts prove out. And ultimate recoveries are high: Moody’s project finance studies put average recovery near 77 cents in the dollar, and its infrastructure default series now runs to 2023. The development banks’ GEMs database, refreshed in October 2025 with data to 2024, independently lands at 73 percent on private-borrower loans. Either figure sits far above senior unsecured corporate paper.
When average recovery is that high, mathematically LGD is the variable doing the work, and LGD ranges enormously deal to deal. A bespoke asset in a thin market, one offtaker, a thin security package, can carry exactly the same DSCR as a liquid, re-contractable asset with a deep resale market and a tight covenant suite, and still bury you on loss. Coverage cannot see that difference but “LGD” is waiting to enter the chat.
Let’s throw some numbers at the wall
Consider two deals, identical on everything the model reports:
(Table: the two deals, shown in the original post.)
Same coverage, same modelled PD, and the expected-loss component of the fair margin differs by 35 bps, entirely on recovery. Price both at the same spread and on a $500m facility you leave $1.75m a year on the table on Deal B, every year, for a risk the coverage ratios swore was identical. That is before the heavier capital charge B also needs to be carried.
On the face of it, the two PDs match. Deal B’s single offtaker and merchant tail make its CFADS more volatile, so its real EDF is closer to 1.8%. Its expected loss becomes 90 bps, six times Deal A’s, and the margin you should be charging is 75 bps wider. The model showed you two identical credits. From a risk perspective they were never the same deal.
The bill, and who pays it
The example represents the lender’s bill: you systematically win the deals you underpriced and lose the ones you did not. IMHO this is a savage demonstration of adverse selection.
The borrower’s bill is the mirror image. Present only DSCR and you leave money on the table, because a credible low-LGD story, strong security, a liquid asset, genuine re-contractability, is a pricing argument, and so is a low-EDF story built on hedged revenue and stable CFADS. Every structuring lever you actually care about, reserve accounts, cash sweeps, security, covenants, moves PD and LGD, not just DSCR. Model only coverage and you are tuning the dashboard while ignoring the engine – with my F1 lens applied.
Let’s up our game in PF land
You do not need a maths PhD or a full ratings replica model. A defensible cumulative PD curve and one honest LGD assumption turn three cushion ratios into an expected loss, and an expected loss into a price. That is not academic garnish. That is the number the whole model was supposed to produce.
The coverage suite was never three ratios. It was always five. Ship or rely on models that are missing two of them is an opportunity lost.
In pulling this together, thank you to:
Moody’s Investors Service, Infrastructure Default and Recovery Rates 1983-2023, and Default and Recovery Rates for Project Finance Bank Loans 1983-2020. GEMs Consortium, Default and Recovery Statistics: Private Lending 1994-2024 (EIB and IFC, October 2025), reporting an average annual recovery rate of 72.9 percent and an average default rate of 3.54 percent across 10,000-plus private borrowers.
First published on Nick Crawley’s Substack.
