Home/AI Orchestration · Model Risk/Part I

All series · Episode 00 · SITUATION · Start here

The CEH-001 situation

Bid/ask/mid, then CEH-001. Credit spreads in bps, peer Recipe A (~1.50) vs curve Recipe B (~2.10). Midpoint 1.80 denied. Read before Ep01.

Ep00 · CEH-001 Stock-trading desk 1.50 / 2.10 · deny 1.80

Why start here

CEH-001 is this series's running exam case: a credit-equity hybrid on an equity desk. It is not a vanilla CDS and not a listed share. Two honest marks land near 1.50 and 2.10. The midpoint 1.80 is denied. Later episodes assume you can see that fight. If you skip this page, "refuse 1.80" will sound like taste instead of risk.

No Clearance classes yet. No heavy math. One morning, two rulers, one committee that wants peace, and software that will invent that peace unless someone stops it.

Bid, ask, mid (retail screen)

Bid: highest price a buyer offers now. Ask (offer): lowest price a seller accepts. The gap is the bid-ask spread. Mid is usually (bid + ask) / 2.

Toy: bid 99, ask 101, mid 100. On a liquid stock the crowd prints bids and asks all day. Your mark can often track mid. You did not invent the price.

Keep three words separate:

Notional is position size. Mark is what that size is worth under today's recipe. The CEH-001 fight is about the mark.

What "no unique mid" means

It does not mean "finance is mystical." It means there is no public two-sided market that quotes bid and ask for this exact claim every morning, so there is no market mid the firm can treat as the day's truth. Someone must build a model-based mark from other markets that do print.

CEH-001 is a bespoke package: credit features plus equity-linked features (barriers, conversion-style rules, call windows, recovery on default). A plain CDS is mostly default insurance. CEH-001 is hybrid, so both a spread-style mark and a cashflow/discount mark can honestly disagree.

Teaching sketch (fictional): while Company ABC stays healthy, pay coupons on schedule. If ABC's stock breaches a barrier for enough days, an equity-linked settlement changes the remaining payoff. On default, recovery rules replace the schedule. Early call windows may exist.

That package may never trade publicly. Then: no bid, no ask, no exchange quote for CEH-001 itself. The desk must estimate value.

Bridge you already know. A thin listed stock with bid 10 and ask 14 has mid 12, but you cannot trade size at 12. Two honest marks can still disagree. CEH-001 is the extreme: the bid-ask ladder for that instrument is missing. Liquid stocks are city houses with sale comps. CEH-001 is a one-off mansion priced from nearby streets (proxies).

What this series is not

Who is in the room

An equity research and trading desk with ordinary shares, options where needed, and a thin sleeve of structured single-name exposures the firm cannot exit on a liquid print. Roles that matter this morning: desk lead (risk file + client call), quant engineer (feature pack / as-of), model owner (scoring crowd), risk co-owner (can refuse alone), an agent that drafts tickets, and a committee slide that demands a single number.

Hold this: the fight is not predicting tomorrow's stock for sport. It is what number enters the books for a position that will not print its own truth.

Novice kit: credit spread, basis points, peers

Read this once before Recipe A. It is the vocabulary most people trip on.

Credit spread is the extra money investors demand to take on a company's credit risk, versus safer money. Wider spread means markets want more pay for the same risk; tighter means less fear in the price.

Basis point (bp) is how spreads are quoted:

So if peers trade around 200-240 bp and our issuer is at 300 bp, the market is saying this name is meaningfully riskier than the peer set.

NameCredit spread
Peer A200 bp
Peer B220 bp
Peer C240 bp
Our issuer300 bp

Why Recipe A uses peers instead of only the issuer's cash-flow sheet:

Peers are not CEH-001. They are sensors. The frame card must name which peers are allowed, or "the mark" becomes an undeclared mix of whatever was convenient that morning.

Two recipes for the same rare claim

Estimating CEH-001 is like appraising a rare house nobody sells every day. Two honest teams, two methods, two numbers. That is the story.

Recipe A - spread frame: "how risky versus peers?"

Question: how much extra do investors demand for this issuer's distress, relative to similar names? That extra is the credit spread (bps). A contracted map turns "spread world" into the desk's value units for this product.

On this morning, after the peer table and the near-optimal crowd inside Recipe A, the centre lands near 1.50, band [1.42, 1.58].

What is 1.50? An internal value unit on the ticket (points, percent of notional, or another model unit). Episode 00 keeps the unit abstract so you watch the fight between frames. When inventory is 40,000 units, a 0.60 gap becomes money.

What is the band? Different allowed runs of Recipe A give slightly different values. Centre near 1.50; plausible range [1.42, 1.58]. Like a weather forecast: expected 30°C, likely 28-32, not fake precision.

Recipe A exists because some consumers (limits, stress packs, hedge overlays) already speak spread language. Forcing them onto pure cashflow present value makes them translate silently and lose the audit trail.

Recipe B - curve frame: "what cash does the contract pay?"

Question: what cash will this contract pay under the legal schedule, and what is that worth today?

Toy schedule (teaching only):

WhenPromised cash (toy)
Year 1100
Year 2100
Year 3100
Final1000

You do not add 100+100+100+1000 and call it today's value. Later money is usually worth less than money now. Turning future amounts into today's money is present value (PV), done with a discount curve: discount factors by maturity, built from liquid rates instruments (deposits, futures, swaps) under named day-count and interpolation rules. The equity desk reuses that infrastructure the way it reuses a calendar. Silent convention swaps change the mark without changing coupon dates.

Pipeline in one line: contract cashflows, then discount curve, then default/recovery assumptions, then PV, then the same value units as Recipe A.

On this morning Recipe B lands near 2.10, band [2.01, 2.22].

Recipe B exists because other consumers (client fair-value language, some accounting paths, cashflow-bucket hedges) speak schedule language.

1.50 versus 2.10

Gap: 2.10 - 1.50 = 0.60. Same issuer, same legal document, same clock, two contracted questions. The book argument is not primarily "will the stock rise tomorrow?" It is: what value enters the books today?

One sentence. Recipe A values CEH-001 from market-implied credit spreads of issuer and peers and lands near 1.50. Recipe B values contractual cashflows with discount curves and default assumptions and lands near 2.10. The series exists so software cannot average those into 1.80 and call the fight settled.

Mark, quote, prediction

Quote: what the market advertises now (bid/ask). Mark: what the firm books. Prediction: a view about a future print. CEH-001's morning is a marking problem. Averaging two marks into 1.80 is a marking cheat, even if someone renames it "forecast consensus."

Proxies, stock, and the curve (short)

When the claim does not print, the desk borrows truth from markets that do: issuer bonds if any, credit quotes, equity and options, peers, sector instruments. None of those is CEH-001. Each is a signal under a named card.

The issuer's equity is another public sensor of the same firm. Equity selloffs often travel with wider credit spreads; rallies with tighter ones. The link is a shared narrative channel, not a law that stock price equals the hybrid mark.

Discount curve: priced calendar for money. Credit on top of discounting (hazard, recovery, or safe curve plus spread) is how Recipe B turns a schedule into a present value. When Part II talks about day-count swaps or tomorrow's ten-year already in today's table, that is time-infrastructure dishonesty poisoning an equity mark, not a rates textbook hijack.

Both can be right; the midpoint is still wrong

1.80 is the average of 1.50 and 2.10. In a meeting it feels fair. It is a third geometry nobody contracted. It sits in neither band. The corridor between 1.58 and 2.01 is empty of declared support.

Inventory of 40,000 units:

Mark usedPosition valueDistance from spread anchorDistance from curve anchor
1.50$60,0000$24,000
2.10$84,000$24,0000
1.80$72,000$12,000$12,000

The midpoint did not shrink the ambiguity. It hid $12,000 of error in an unknown direction behind a single float that every downstream system will treat as a measurement.

You are not fighting elegance. You are fighting a $12,000 invisible error wearing the costume of a single number.

Two jobs, one forbidden compromise

1.80 tries to serve all three by serving none honestly. That is why deny_midpoint is operational, not philosophy: job identity is part of the mark.

Honest call: "Under the spread frame, near 1.50 with band [1.42, 1.58]. Under the curve frame, near 2.10 with band [2.01, 2.22]. We are not averaging those. For your schedule-based mandate we use the curve frame today, and that choice is logged."

Dishonest dashboard: Consensus Mark 1.80, green checks, agent summary "models agree ~1.80," promote ticket half-filled. All of that can be generated from the same dual clouds if the pipeline was never told to refuse.

One morning, condensed

09:05-09:40. Peer spreads move. Curve points update. Freeze the information pack at the decision clock. Tomorrow's ten-year already in today's table cheats the exam.

09:47. Dual marks on the ticket: 1.50 and 2.10, with bands and frame ids. No blend field.

10:06. Bad agent writes mark = 1.80 from a "consensus" tool. Good agent cannot: schema and harness refuse it.

10:20-10:24. Frame honesty needs both marks. Leakage checks need a legal pack. Hard fail: no spend lease. Soft research rights still keep both clouds visible.

11:30. Client call chooses a frame on the record, or keeps dual marks. Not a revelation of a single true float.

The fight is not "which AI is smarter." It is whether the firm preserves a real ambiguity or launders it into a float.

Where AI helps, and where it peacemakes

Later episodes are the refusal machinery around those costumes: frames, as-of honesty, powered eval, Clearance rights, book-level stress.

Numbers to keep

ObjectValueMeaning
Spread-frame mark1.50Peer-spread geometry
Spread band[1.42, 1.58]Stakeable support under that frame
Curve-frame mark2.10Cashflow-discount geometry
Curve band[2.01, 2.22]Stakeable support under that frame
Frame gap0.60Honest disagreement width
Denied midpoint1.80Uncontracted third geometry
Inventory toy40,000 unitsMakes a 0.60 gap worth real money

These return for the whole series. Learn them as a case.

How to read what comes next

Whenever a later chapter says "refuse 1.80," return here: two panes, two jobs, one empty corridor, a committee that wants peace, and a position whose honest statement is a pair.

Next. Open Ep01: Bertrand's Paradox and AI model risk. Part I index.