What Sits Beneath the Model: The Data Layer Institutional Decisions Actually Run On

Two models for the same building, built by two analysts in the same firm, disagree on the vacancy rate. One says 7.8%. The other says 9.1%. Both are current, both name a source, and when the difference is noticed in committee, nobody in the room can say why they differ without leaving to make a phone call.

The explanation, once someone makes the call, is that one analyst took the figure from a broker’s quarterly survey published in March and never revised, and the other pulled it from the firm’s database subscription on the morning the memo was finalised, by which time the vendor had restated the quarter twice as late lettings were captured. Both figures were true on some date, and both use the same stock definition. Neither model records which date, and neither records what kind of source it was reading: an opinion published once, or a database that keeps moving after it has been read. Both could carry a perfect citation and the committee would still not know that. It is a property of the source, not of the citation, and the workbook has no field for it. The model can compute anything the analyst asks of it. It cannot say what it does not know.

This is a data-architecture problem rather than an analyst error. A model sits downstream of a layer of data it does not describe, and it inherits two properties from that layer: how accurate the figures are, and whether anyone outside the deal team could confirm them. Firms treat these as one property. They are two, and they rank sources differently. An earlier note dealt with the anatomy of a single cited figure; this one is about the stratum the citation points into, and what a firm can do about the fact that most of its numbers come from the middle one.

The middle stratum exists because, as the RICS professional standard on comparable evidence puts it, full details of sales and lettings are rarely available to the public even in the most transparent markets. Somebody assembles what is missing, and the assembler is a broker with a position or a vendor with a subscription. Even the benchmark index is downstream of other firms’ systems: MSCI’s methodology describes valuation data being read into owners’ and managing agents’ in-house systems before it reaches the index. A market fact has usually passed through several maintained databases on its way to the committee, none of which recorded the passage.

The Smart Bricks Research Lab has been working through where institutional inputs originate, drawing on published registry statistics, the RICS and MSCI methodologies and the process framework behind its own screening. What follows is a taxonomy, the published measurements that exist for each layer, and an architecture.

Beneath every institutional model sit three kinds of market data.

The opinion stratum. Broker quotes, agent surveys, quarterly market reports, verbal guidance on where yields are trading. Produced by a party with a position, revised without announcement, and carrying no published error rate. Estimated rental values and submarket vacancy figures usually begin life here.

The database stratum. Proprietary and vendor databases, appraisal-based indices, transaction databases assembled from voluntary contributions, and the firm’s own property-management, accounting and modelling systems. Produced by a vendor or by the firm itself, revised as late records arrive or as appraisals update, and disclaiming rather than measuring their own revision.

The registered stratum. The public record: a title lodged with a land registry, a transaction recorded against it, a filed set of accounts, a caveat or a stamp-duty return. Produced by a registrar after a counterparty accepted it, corrected through a documented process, and publishing what it rejects and, where a market insures title, what it misses.

The taxonomy classifies data about the market. It does not classify the documents a firm holds as a party to them: the executed lease with its amendments, the facility agreement, the bank statements showing what was collected. Those are the evidence from which the firm’s own database-stratum figures are derived, and the evidence Section VII reconciles them against.

The RICS standard on comparable evidence already treats its middle category this way: Category B covers published sources and commercial databases, weighted by relevance, authority and verifiability. What that weighting captures, and a typical model does not, is that reliability and verifiability are separate questions. Of the seven attributes the standard lists for good evidence, a well-maintained proprietary database can score highly on five and still fail two structurally, recent and verifiable, because its contents are the vendor’s records as they stood on a date, and a stranger can only check them by holding the same subscription and looking on the same day.

The registered stratum has the opposite profile. It is often less complete than a good database, slower to reflect a transaction, and residential-weighted in many jurisdictions. Its distinguishing property is that anyone with the title number can look. In England and Wales, the register holds more than 27 million titles, covers more than 90% of the land mass, and the title it records is guaranteed by the government; Price Paid Data drawn from it is released monthly under an open licence at no charge. That is verifiability by a stranger built into the source rather than added later.

Stratum	Who produces it	How it revises	What it publishes about itself	Checkable by a stranger	Typical inputs
Opinion and broker survey	Who produces itProduced by a party with a position	How it revisesRevises silently	What it publishes about itselfNo published error rate	Checkable by a strangerNo	Typical inputsQuoted market rent, submarket vacancy, yield guidance
Proprietary or vendor database	Who produces itProduced by a vendor or the firm's own systems	How it revisesRevises as late records or new appraisals arrive	What it publishes about itselfDisclaims revision, does not measure it	Checkable by a strangerOnly with the same subscription, as of today	Typical inputsAppraisal-based index, transaction database, the firm's rent roll
Registered public record	Who produces itProduced by a registrar after a counterparty accepted it	How it revisesCorrected through a documented process	What it publishes about itselfPublishes its requisition rate; where insured, its insured losses	Checkable by a strangerYes, by title number or record ID	Typical inputsLand Registry title, Price Paid Data entry, stamp-duty return, lodged caveat

Reliability and verifiability are different axes. The middle stratum can be highly reliable and still uncheckable by a stranger, and it revises without recording that it did. Executed documents the firm holds (leases, loan agreements, bank statements) are not a stratum of market data; they are the evidence the database stratum is derived from and reconciled against.

Source: RICS, Comparable evidence in real estate valuation (2019, reissued 2023); HM Land Registry; Urban Redevelopment Authority; MSCI Property Indexes Methodology (July 2026)

The defining property of the database stratum is that it changes after the fact, and it changes in three distinct ways that get lumped together as “data quality.”

The first is late capture. A transaction database is only as complete as the deals reported to it by publication; deals closed in a quarter keep surfacing for months afterwards, and the quarterly total moves accordingly. The publishers say so themselves: MSCI’s capital-trends publications carry a standing note that the data is subject to future revision, standard language for any data vendor and, here, an accurate description of the mechanism. A volume figure quoted in a memo the month it was published is provisional by the publisher’s own account, and the memo will not be updated when it settles.

The second is appraisal smoothing. Valuation-based indices are built from appraisals, and appraisers anchor on prior values and on stale comparables. Published research going back to Fisher, Geltner and Webb in 1994 established that appraisal-based indices are smoothed and lag transaction prices, and a literature of unsmoothing filters grew up to correct for it. The lag is established; its size is contested, and Bond, Hwang and Marcato found that the common aggregate filters overstate smoothing at the individual property level. For a committee, an index-derived capital growth figure is not wrong so much as late, and the correction applied to it depends on which academic camp it came from.

The third is valuation cadence. MSCI’s methodology describes the indices as measuring unlevered, valuation-based performance, and it requires that a constituent portfolio be valued at least once during the measurement period, with capital values otherwise rolled forward and adjusted for capital expenditure. Under that rule a monthly or quarterly index observation for a given asset may rest on a valuation some months old, carried forward arithmetically: a maintained view, not an observation.

None of this is a criticism of the publishers, who document their methods more carefully than most subscribers read them. The point is structural. A number from this stratum is a photograph of a database on a date. Next quarter the database will have moved, and nothing in the model that consumed it will know.

The argument for the registered stratum is counter-intuitive. It is the strongest layer beneath the model on the axis that matters to a stranger, because the registrar counts and publishes what it sends back on the way in, and, where a market insures title instead of registering it, the insurer counts and publishes what the record missed on the way out. Neither count is an error rate for the accepted record; registered is not the same as correct, and Section IX returns to that. But no other stratum publishes anything comparable about itself.

HM Land Registry publishes the rate at which the applications it receives are sent back for defects the applicant could have avoided. Across professional customers with ten or more completed applications, over a six-month period the registry analysed and reported in November 2025, the average avoidable-requisition rate was 4.6%. One customer in five ran below 1%. Two hundred and ninety-four customers ran at 10% or more. The registry reported issuing more than 100,000 fewer requisition letters in 2025-26. That is a sent-back rate on the register’s intake, published by the registrar, by customer, with a trend.

The United States records deeds county by county and guarantees no title, and what its market does instead is instructive: it insures the record and counts the losses. The American Land Title Association reports that the title industry wrote USD 18.5 billion of premium in 2025, up 13.8% on the prior year, and paid USD 667 million in claims, against USD 676 million in 2024. Our own arithmetic puts claims at roughly 3.6% of premium; ALTA does not state a loss ratio and the ratio should not be attributed to it. The more useful number is in ALTA’s claims analysis with Milliman. In the 2025 edition, which covers residential transactions, approximately 40% of losses and defence costs on refinance transactions arose from fraud and forgery, defects that cannot be identified by searching the public record. Where the record is insured rather than guaranteed, it fails in a known direction, at a measured rate, and the party carrying the loss publishes the analysis.

Coverage is disclosed too, where a registry is run well. The Urban Redevelopment Authority in Singapore notes that caveat lodgement is voluntary, that caveats historically captured about 80% to 90% of sub-sale and resale transactions, and that since 2015 those figures have been compiled from stamp-duty records instead. A database that discloses neither its coverage nor its revision policy is one the committee weights by habit.

• 4.6%

  Average avoidable-requisition rate across HM Land Registry professional customers with ten or more applications, six-month period reported November 2025

  HM Land Registry

• 27m titles, 90% of land

  Register coverage in England and Wales, title guaranteed by the government

  HM Land Registry

• USD 667m

  US title insurance claims paid in 2025, against USD 18.5bn of premium, in a market that records deeds without guaranteeing title

  ALTA, May 2026

• ~40%

  Share of residential refinance title losses arising from fraud and forgery that no public-record search could find

  ALTA and Milliman, November 2025

• 80% to 90%

  Share of Singapore resales captured by voluntary caveats before 2015, since replaced by stamp-duty records

  URA via data.gov.sg

The strongest stratum is the one whose defects are counted, by the registrar on the way in and, where title is insured rather than registered, by the insurer on the way out. No equivalent is published for a broker survey or an appraisal index.

Source: HM Land Registry (GOV.UK and HM Land Registry blog); ALTA press release, 22 May 2026; ALTA and Milliman, 2025 Analysis of Claims (November 2025); URA dataset notes on data.gov.sg

The model will keep inheriting whatever the layer beneath it knew. The choice is whether the firm knows what that was.

Smart Bricks Research Lab

Turn the taxonomy inward and it becomes uncomfortable. The rent roll in the property-management system, the model in the modelling tool and the ledger in the accounting system are proprietary databases in every sense that matters. They are maintained views of the executed documents the firm holds, kept current by people, and they carry the same two defects as any vendor database: they revise without recording it, and they attach figures to identities that were never resolved.

The rent roll is the clearest case. It should be a derived view, computed from the executed leases and their amendments with each cell pointing back at the clause it came from. In practice it is usually a maintained spreadsheet or a system table keyed in by hand, and a maintained rent roll drifts from the leases. In the taxonomy above the rent roll is a database-stratum figure inside the firm’s own walls, one step removed from the documents it summarises, and the model treats it as though it were registered.

The identity problem is the one nobody budgets for. Real estate has no universal entity identifier. Tenants are recorded under trading names, legal names and abbreviations; the same building carries a different key in the property system, the accounting system and the debt schedule. The costliest version is the confusion between the entity that signed the lease and the entity that guarantees it: link the signatory to the wrong parent and every concentration report downstream is summing exposure across a company that never put its name to the document. Resolution cannot be a periodic clean-up, because the counterparties keep changing.

Great Britain shows that a property key is achievable, and where it stops. The Unique Property Reference Number is a persistent identifier allocated primarily by local authorities, carried by roughly 40 million addressable locations, published by Ordnance Survey as open data and mandated as the government’s open standard for property and street information. A firm can key every British asset to it; its assets in the Netherlands, the Nordic countries or the United States carry different keys under different authorities, and the firm’s own systems usually carry none of them.

The standards a data team will reach for divide neatly by what they cover.

OSCRE’s Industry Data Model defines the entities and relationships of a real estate dataset and publishes them as schemas. INREV publishes a Standard Data Delivery Sheet for the exchange of fund data between managers and investors, with an ESG version since November 2023. Outside real estate, ISO/IEC 25012 defines the characteristics of data quality, among them accuracy, completeness, consistency, credibility and currentness. The vocabulary exists, and has for years.

What none of these documents does is set a tolerance. OSCRE defines the shape of a lease record; it does not say what share of lease records must load cleanly before a migration is accepted. INREV defines the fields an investor should receive; it does not say how many may be blank, or how far the reported figure may sit from the audited one before the sheet is rejected. ISO/IEC 25012 names currentness as a quality; it does not say how old a rent roll may be before it stops being current. The consequence shows up at every system go-live: trial balances, unit counts, lease counts and receivables tied out by hand against the legacy system, to a tolerance nobody wrote down, because no standard supplied one.

The schemas say what a well-formed figure looks like; none says whether a given figure arrived intact, from where, or as of when. A firm that wants those properties has to add a layer that supplies them, and carry its own instruments for it. The next section names both.

The architecture that follows is modest, and it is better stated as a set of tests than as a design, because a committee will meet it as a proposal, from its own data team or from a vendor, and needs to know what to require. The ledger, the property system and the model remain the systems of record. Nothing is migrated and nothing is re-keyed. Above them sits a layer required to do five things those systems were never designed to do.

It reads without moving. Leases, amendments, the rent roll, the general ledger, the debt schedule and the underwriting model are read where they live. The test is whether it could be removed without the systems beneath it noticing; a layer that fails it has become another system of record, with the revision history of one.

It stamps every figure. Each value the layer touches carries the stratum it came from, which is to say opinion, database or registered record; the date on which it was true; and the identity it attaches to. The stamp adds stratum and identity to the five parts of a cited number set out in the earlier note; it does not replace them. A committee that can see the stratum can weight the figure as RICS asks a valuer to weight Category B evidence.

It resolves identity. The property, the tenant, the guarantor and the fund behind every stamped figure are matched to one governed record, and a subsidiary is linked to its parent before any exposure is summed. The test for the layer is whether it carries its queue of unresolved matches openly, with a named owner, or hides it inside a matching rule nobody has read.

It reconciles the database against the document and the register. This is the test the systems of record cannot run on themselves, and it has to be specific about what each source can corroborate. The executed lease and its amendments confirm term and rent as agreed; bank receipts confirm what was collected; the registered title confirms who owns the asset and what encumbers it, and in England and Wales a lease reaches the register only if its term is long enough to be registrable, while the title carries no rent. For most commercial income, then, the rent roll is tested against the lease and the bank, with the register confirming that the asset and its owner are what the rent roll says they are. Where the sources disagree, the disagreement is surfaced with its evidence rather than averaged away, every reporting cycle rather than once at migration.

It proposes, and a person approves. Corrections to the rent roll, updated assumptions, a re-linked tenant identity: the layer drafts them, attaches the evidence, and writes nothing to the system of record until someone with authority signs off. It is the same control institutional capital would require of any analyst who found a discrepancy.

Since no external standard exists for the layer to report against, it carries three instruments of its own, and a committee should ask to see them: the identity match rate and the age of the oldest unresolved match; the break rate per reporting cycle, meaning how many rent roll lines disagreed with the lease or the bank beyond the firm’s written tolerance, and how many were cleared; and the share of memo inputs by stratum, which the next section shows a CIO how to count by hand. A layer that cannot report all three is asserting that it works rather than showing it.

Test	Systems it touches	Instrument that shows it is met	What failing looks like
Reads without moving	Systems it touchesProperty system, ledger, debt schedule, model, document store	Instrument that shows it is metThe layer could be removed without any system beneath it noticing	What failing looks likeData copied and re-keyed into a new store that now needs its own reconciliation
Stamps every figure	Systems it touchesEvery value drawn from those systems	Instrument that shows it is metStratum, as-at date and identity visible on each figure	What failing looks likeA figure with a source name but no stratum and no date
Resolves identity	Systems it touchesTenant, guarantor, property and fund records across systems	Instrument that shows it is metMatch rate, and age of the oldest unresolved match, reported each cycle	What failing looks likeA matching rule nobody has read and a queue nobody owns
Reconciles database against document and register	Systems it touchesRent roll against executed lease and bank receipts; title for ownership and encumbrances; filed accounts for covenant	Instrument that shows it is metBreak rate per cycle against a written tolerance, with lineage to the source	What failing looks likeDisagreements averaged, or checked once at migration and never again
Proposes, then a person approves	Systems it touchesCorrections and re-linked identities bound for the system of record	Instrument that shows it is metEvery write-back carries its evidence and a named approver	What failing looks likeAn automated process mutating the ledger

The ledger, the property system and the model stay authoritative. Each test names the systems it touches, the instrument that shows it is being met, and the failure a committee should recognise.

Source: Smart Bricks Research Lab

The argument above can be tested in a week without buying anything, and the test differs from a citation check. A citation check asks whether each number names a document. A stratum audit asks what kind of thing the document was.

Take the most recent investment committee memo and list its ten most material inputs: the passing rent, the vacancy figure, the market rent, the rent growth assumption, the exit yield, the comparable transactions, the capital expenditure allowance, the operating cost base, the debt terms and the purchase price. For each, the six questions in Exhibit 4 are answered in writing, by the analyst who used the input, before the answers are compared.

For the passing rent the answers usually run as follows. Stratum: database, because the figure came from the rent roll in the property system, with the executed lease sitting in the data room unreconciled against the line. Date: the rent roll’s export date, rarely the memo date. Identity: the tenant’s trading name as keyed by whoever set up the lease, which may not match the ledger. Registered check: counsel read the title, but the title number was never joined to the rent roll or the model. Revision sensitivity: the figure will move if an amendment surfaces, and nobody has asked by how much. Opinion label: not applicable, since passing rent derives from an executed document. One input, six answers, and the memo’s largest single number turns out to be a database photograph with an unresolved identity, no as-at date, and a document behind it that nobody tied to the figure.

Expect most inputs to sit in the middle. Market rent and vacancy will come from a broker survey or an index. Comparables will come from a subscription database. The registered record will typically have been consulted exactly once, by counsel at legal review, and its contents will not appear in the model at all. None of that is a scandal. The count of middle-stratum inputs is the finding, and it is the number a CIO should want to see fall over time.

Three inputs deserve separate treatment. Estimated rental value, rent growth and exit yield have no registered stratum and never will. They are opinions by construction, and the right discipline is to label them as such, with the person who chose them and the evidence they were looking at, rather than to dress them as data by citing the survey they were lifted from. A memo that marks its opinions as opinions is more defensible than one that gives every input the same typeface.

1. 1

   Stratum

   Did this number come from opinion, a database or a registered record, and if a database, which executed document is it a view of?

2. 2

   Date

   On what date was it true, and has the source been revised since?

3. 3

   Identity

   Which identifier does it attach to, and would the ledger find the same tenant or asset by it?

4. 4

   Registered check

   Was a registered record consulted for this fact, and was it joined back to the model?

5. 5

   Revision sensitivity

   If the database figure moved by a revision of the size its publisher warns of, would the recommendation change?

6. 6

   Opinion label

   If the input has no registered stratum (ERV, rent growth, exit yield), is it labelled as an opinion with a named owner?

Run on the ten most material inputs of the last IC memo. The count of inputs that came from a database or an opinion is the finding.

Source: Smart Bricks Research Lab

The registered stratum is cheap or free wherever it exists, and its coverage is jurisdictional and partial. Price Paid Data in England and Wales is free under the Open Government Licence, and it is also residential by definition, so a commercial buyer gets the government’s guarantee of title for the price of a title search and no registered price series at all. The Unique Property Reference Number stops at the Great Britain border. A firm operating across England and Wales, the Netherlands, the Nordic countries and the United States has a different top stratum in each, and it has to know which one it is standing on before it weights a figure.

Registered is not the same as correct. The 4.6% avoidable-requisition rate is a sent-back rate on the register’s intake, and the register is only as good as the applications it accepts. The ALTA and Milliman analysis shows a stable class of defect, fraud and forgery, that no search of the record will find, which is part of why a market that insures title rather than guaranteeing it carries USD 18.5 billion of annual premium against the record. A firm that treats the registered stratum as the truth has swapped one unexamined assumption for another.

The cost of the layer has a shape, and most of it is people rather than software. The stratum audit in Section VIII is a week of one analyst’s time and produces the baseline count. The tolerance document is the CIO’s to write once and review each year, since no framework will write it: how many records may fail to match, how far the rent roll may diverge from the leases before escalation, which control totals must tie out each cycle and to what precision. The standing cost is the exception queue. Identity resolution never finishes, because tenants restructure, guarantors are replaced and buildings are subdivided; that is reviewer time every cycle, budgeted as a named role, and a firm that cannot name the person who clears the queue has not budgeted for it. Without the tolerance document the layer surfaces exceptions nobody has agreed how to treat, and it becomes the kind of system that is switched off for producing noise.

The layer does not improve the opinion stratum. Broker views, growth assumptions and exit yields remain opinions after they have been stamped; the stamp makes them visible as opinions, not better ones. A firm that expects a data layer to improve judgment will be disappointed; one that expects it to show where judgment has been substituted for record, and by whom, will get what a data layer can give.

Return to the two vacancy rates. Put through the layer, both arrive at the committee stamped: 7.8%, opinion, broker survey, as at March, unrevised; 9.1%, database, vendor subscription, as at the memo date, restated twice since the survey. The committee sees why they differ without the phone call, and it sees that neither is a registered fact, because no registry records vacancy. That is what the layer can give a committee: not better numbers, but numbers that say where they came from.