The inventory and the tiering call

It is common industry practice to maintain a comprehensive set of information for models under development and models in use, and the level of detail may reasonably vary with a model's complexity and usage. That much sounds like a filing exercise. The clause that turns it into a design problem is the next one: an effective inventory holds enough information to understand model risks and support the management of them both at the level of the individual model and in aggregate.

Read that as a requirement on the columns rather than on the rows. Individual-level management is satisfied by a row: who owns this, what is it for, when was it last validated, what is still open. Aggregate management is a question asked across rows, and you can only ask it about things you wrote down. Nobody can see that eleven models share one data feed if the feed was never a field. Nobody can see that a whole portfolio rests on one assumption about bureau comparability if assumptions live in PDFs. A register of names, owners and dates is a list of models. It is not an inventory, because the second half of the job is structurally impossible with it.

• Inventory. The comprehensive record of models under development and in use, detailed enough to support management of model risk individually and in aggregate.
• Exposure. The measurable half of materiality. How much rides on the output: portfolio size, business impact, the volume of decisions the thing touches.
• Purpose. The judged half. What the output is for, with uses that meet a regulatory requirement or manage financial risk weighing more than uses that size a campaign.
• Materiality. Purpose together with exposure. In the current guidance these two are what determine a model's materiality, and overall model risk is inherent risk read in the context of it.
• Tier. An editorial device, not a regulatory object. The guidance asks that the nature and rigour of review align with the model's approach, use and materiality; a numbered tier is one house convention for writing that alignment down.
• Aggregate risk. Interactions and dependencies among models, and reliance on common assumptions, data or methodologies. The reason the columns look the way they do.
• Immaterial. A scored outcome, not an exemption. An immaterial model is still identified, and still monitored for the conditions that would make it material.
• Re-tier. The act a person performs when purpose or exposure moves. No field does it automatically, which is the subject of Part 2.

Aggregate model risk reflects interactions and dependencies among models, and reliance on common assumptions, data or methodologies. Every clause in that sentence describes something that is invisible inside a single model document, because a single model document has no reason to mention what the model next door is doing.

The practical consequence is that the inventory is the only artefact in the whole discipline that can see this. A validation report covers one model. A monitoring pack covers one model. The inventory is where twenty individually reasonable models turn out to share one vendor feed, one macro assumption and one preprocessing library, and where somebody can notice that a correction to any of the three moves the whole book at once. If your columns cannot express that, the aggregate view is not being done anywhere, no matter how many committees discuss it.

Documentation gets treated as a tax, which is why it gets written last and read never. It is worth being specific about the jobs it does, because each one implies a field. It supports continuity of operations, so the model survives the departure of the person who built it. It tracks recommendations, responses and exceptions, so an argument that was settled two years ago stays settled and the reasoning is recoverable. And it supports the management of remediation, so a finding has a state rather than a memory.

Read as column requirements, those three sentences produce most of the right-hand list. Continuity means owner, purpose, feeds and assumptions written where a stranger can find them. Tracking recommendations and exceptions means a findings register that the inventory points at, including the exceptions that were granted and by whom. Remediation management means each open finding carries an owner and a date, which is what separates a condition from a sentence.

The scope covers models under development as well as models in use, and teams routinely read this as pedantry. It is not. A model that is being built is a model that somebody is already planning around, and the gap between "we are building a pricing model" and "the pricing model exists" is precisely the window in which validation has to be scheduled, data access negotiated and a tier assigned. A model that appears in the inventory on the day it goes live has already skipped the part where the review could have changed anything.

There is a second reason, which is that the pipeline is where the aggregate view gets its warning. Three teams independently building on the same new feed is a concentration that has not happened yet, and the cheapest moment to notice it is while all three are still choosing.

The failure pattern is a tiering sheet whose inputs are words: "high impact", "significant", "business critical". Every model owner believes their model is business critical, and they are not being dishonest, they are answering a question that has no shared unit. Replace the words with two recorded quantities and one recorded sentence and the argument gets much shorter.

For exposure, pick a measurable thing the model touches and record the number with its unit and its date: balances scored, decisions per year, provisions influenced, notional covered. Any of these is defensible; none of them is a synonym for the others, so the inventory has to say which one it holds. For purpose, record what the output decides, in a sentence naming the decision and who acts on it, and record whether that use meets a regulatory requirement or manages financial risk, because those uses weigh more. The tier is then a stated function of the two, applied the same way to every model, and if somebody wants a different tier they have to argue with the function rather than with you.

Rigour is commensurate with risk, which means some models correctly get a paragraph rather than three weeks. The part that gets dropped is the second half: a model judged immaterial is still identified, and still monitored for the conditions under which it would stop being immaterial. Those conditions are the useful artefact, and they have to be written at tiering time, while the reasoning is fresh, not reconstructed afterwards by someone explaining a loss.

A condition is specific and checkable. "Re-tier if this model is extended beyond the one product line." "Re-tier if the balances it touches pass the threshold in the tiering function." "Re-tier if its output starts feeding a price rather than a ranking." Each of those is a sentence somebody can be given ownership of, and each of them is a column in the inventory rather than a paragraph in a PDF, because a paragraph in a PDF cannot be reported on.

Exposure changes loudly. Balances grow, volumes appear in a monthly pack, and somebody somewhere is already reporting the number, so a threshold crossing can be automated. Purpose changes quietly, because a change of purpose usually looks like a good idea implemented by a competent person: the scorecard ranks well, pricing needs a risk input, the score is right there, and the integration takes an afternoon. No document is amended, because no document was wrong. The model still does what it always did.

What changed is what the output decides, which is one of the two inputs to materiality. This is also the seam with model use, which session 2 covers: using a model beyond its intended purpose introduces additional uncertainty and risk, and when a use is extended the sound practice is additional analysis of the new usage and its limitations, plus a review of the existing controls. For tonight the inventory consequence is narrow and practical. Put a "purpose last confirmed" date next to the purpose sentence, and make confirming it an act somebody performs, because an undated purpose field ages into fiction without ever being edited.