The credit desk: oversight, alpha & controls

Every PM is handed a mandate — a rulebook of what they may hold: eligible ratings and sectors, single-issuer and structured-product caps, duration and spread-duration ceilings, liquidity floors, prohibited names. Oversight is the machinery that checks the live book against that rulebook continuously — pre-trade (block the ticket) and post-trade (flag the drift). A breach freezes new risk, notifies the PM and starts a remediation clock.

The dashboard below runs eight limits on a real book with CLOs. It opens in breach — the three CLO tranches are one manager (Octagon), tripping the single-issuer cap — with CLO% and illiquidity sitting on watch. Resize any line and the limits recompute live.

When a PM makes 12%, the allocator's first question is: how much of that was skill, and how much was just being long a market that went up? You answer it by regressing the PM's return series on a benchmark. The slope is beta(market exposure you could buy in an index for a few basis points) and the intercept is alpha — the return left over. That residual, month after month, is the alpha series: the real product.

But a few good months can be luck. So you don't just measure alpha — you test whether it's statistically real(a t-stat above ~2) and how efficiently it's earned (the Information Ratio= annualised alpha ÷ tracking error). Below, dial in a PM's true skill and beta, draw a 36-month sample, and watch the analytics try to recover it — note how thin skill hides inside noise until you have enough history.

Oversight and analytics are only as good as the numbers underneath them. Controls are what let the firm — and its auditors, prime brokers and investors — trust and trace every mark. For an illiquid credit book (CLO tranches barely trade), this is where the real risk of misstatement lives.

A CLO pools several hundred leveraged loans and sells the cashflows in tranches, from AAA down to the first-loss equity. The whole game is subordination: losses hit the bottom first, so a junior tranche has to be wiped out before the one above it takes a cent. Credit-risk analytics is about knowing exactly where a given level of defaults lands in that stack — and whether the deal's protective tests still hold.

The lab drives the collateral with a constant default rate, recovery and life, plus a jump-to-default shock. Watch losses climb the structure, the over-collateralisation (OC) tests trip, and the AAA sit comfortably above even severe stress — which is exactly why the mandate above still rates it AAA.

Every model so far assumed we can measurethe risk factors. For a CLO book, many of the factors that move your P&L barely trade — so you can't build a trustworthy statistical model of them. Under the regulatory test (FRTB's Risk Factor Eligibility Test), a factor is modellable only if it has enough real price observations — roughly ≥24 a year with no gap over a month, or ≥100 a year. Everything else is a Non-Modellable Risk Factor (NMRF), and it gets a punitive, poorly-diversified capital charge instead of going into your VaR/ES engine.

The lab lists the factors hitting this book. The liquid ones — rates, IG/HY index spread — pass. Almost everything specific to the CLOs and single names fails: mezzanine and equity tranche spreads, default correlation (which is unobservable, not merely illiquid), recovery rates, reinvestment/prepayment, the CDS–bond basis and long-dated curve points. Together they dominate the capital number.