Data Science Consulting Services: Real Estate Case

Executive Summary

Data science consulting services were engaged by a commercial real estate group in North Texas that could describe every asset it owned but could not answer a portfolio question without three days of manual work. Lease economics sat in scanned documents. Tenant records lived in a customer relationship system that did not agree with the property management system.

The group holds office, industrial and retail assets across several submarkets. Asset managers made renewal and capital decisions on judgement, because the numbers needed to do otherwise were not assembled anywhere.

YlogX built a governed data layer joining property, tenant and lease records, then extracted lease economics from documents into structured fields.

On that foundation sit three decision products. A renewal propensity score, a rent position benchmark against comparable space, and a capital prioritisation ranking.

Projected outcomes include an estimated 25 to 30 percent reduction in time to produce a quarterly portfolio review and an estimated 10 to 15 percent improvement in renewal capture on assets where the score is acted on twelve months ahead. Both are modeled, not measured.

Spending days assembling a portfolio review by hand? Talk to us.

Client Overview

The client is a privately held commercial real estate group operating across North Texas and two neighbouring states. Its portfolio mixes suburban office, light industrial and neighbourhood retail, with a small development pipeline.

The team is lean. A handful of asset managers cover the whole portfolio, supported by a property accounting function and an outsourced facilities operation.

Systems accumulated over two decades. A property management and accounting platform holds the rent roll. A customer relationship system holds tenant contacts and deal pipeline. Lease documents sit in a document store.

Market data arrives from a commercial subscription feed, exported to spreadsheets when somebody needs a comparison.

Nothing about this is unusual for a group of this size. The systems each work. What had never existed was any reliable way to join them into a single view of an asset.

Business Challenges

Every asset manager knew the portfolio well. The difficulty was that this knowledge lived in individual heads and in individual spreadsheets, and it could not be aggregated, tested or handed over.

Four problems compounded, and the first one made the rest structurally impossible to solve.

Why could the group not answer basic portfolio questions?

Because there was no reliable join between systems. The accounting platform identified tenants by an internal code, the customer relationship system by company name typed at deal stage.

The same operating company appeared under four spellings across the portfolio, and two of its subsidiaries looked like unrelated tenants entirely.

Space was the other problem. A suite split between two tenants existed as one record in one system and two in another, so occupancy calculations disagreed depending on the source.

A question as simple as total exposure to a single tenant covenant across all assets could not be answered without a person opening several systems and making judgement calls.

What did the lease documents actually hide?

The economics. The rent roll held base rent, but escalation schedules, renewal options, expansion rights, recovery structures and free rent periods sat only in the executed documents.

Those documents are scanned, often with amendments and side letters filed separately from the original lease.

So any forward looking calculation required somebody to read the lease. Weighted average lease term, rollover exposure and option risk were all manual exercises.

One consequence was that renewal options were occasionally missed until a tenant exercised one, at a rent struck years earlier and well below current market.

Why were renewal decisions made so late?

Because the trigger was the expiry date rather than any assessment of risk. Conversations typically began around six months out, which is late for space that needs re letting.

Nobody had a ranked view of which tenants were most likely to leave, so attention went to whichever expiry was nearest rather than to whichever mattered most.

Signals that would have predicted departure existed but were scattered. Late payments sat in accounting, service complaints in the facilities system, headcount changes in nobody's system at all.

The cost of a late start is concrete. Downtime, tenant improvement spend and leasing commissions all rise when a renewal fails and space has to be re let.

How was capital being prioritised?

Largely by whoever argued most persuasively in the quarterly meeting, informed by real knowledge but not by any consistent comparison.

Roof and mechanical replacement competed with lobby refurbishment and speculative suite fit out, with no shared basis for ranking them.

Deferred maintenance was tracked in a facilities spreadsheet that did not connect to lease exposure, so spend was not weighted by what it protected.

The group suspected it was underspending on assets with concentrated rollover and overspending on stabilised ones. It had no way to test that suspicion.

Renewal options buried in scanned leases? Read more about our approach at data science and machine learning capability.

Business Objectives

The brief was decision support, not reporting. The group had reports. What it lacked was any ranked answer to the question of where attention should go next.

What did the asset management team need first?

A single reliable view of property, space, tenant and lease that every number could be traced back to. Without that nothing else was worth building.

Lease economics as structured fields rather than as documents, so that forward looking calculations could run without anyone reading a contract.

Early warning on renewals, far enough ahead that a conversation could change the outcome rather than merely record it.

And a consistent basis for comparing capital requests, so that the quarterly meeting argued about priorities rather than about numbers.

Why was explainability made a hard requirement?

Because asset managers will not act on a score they cannot argue with. A ranked list without reasons gets ignored on the first occasion it disagrees with experience.

Every renewal propensity score therefore carries its main contributing factors, expressed in terms the team already uses.

That constraint shaped model selection. Interpretable methods were preferred, and where a more complex model was used, its outputs are explained per record.

The requirement also protected the group commercially, since a renewal conversation informed by a score has to be defensible to a broker and a tenant.

What defined success for this engagement?

Portfolio questions answerable in minutes rather than days, with a traceable path from any figure back to its source record.

Renewal risk visible twelve months ahead of expiry rather than six, with reasons attached.

Capital requests ranked on a shared basis that weights lease exposure, not just asset condition.

And the team using the tools without assistance after handover, which was tested rather than assumed at the end of the engagement.

Success was also defined as no new reporting tool. The group had bought analytics software before and had not adopted it, and that history shaped the brief.

Want renewal risk visible a year ahead? Start a conversation.

Solution Strategy

Entity resolution came first and consumed more of the engagement than anything else. Modelling was straightforward once the underlying records agreed with one another.

Delivery ran on a YlogX solution accelerator for the pipeline and monitoring scaffolding, adapted to a real estate entity model.

How do data science consulting services approach entity resolution?

By defining the entity model before touching any source. Property, space, tenant, lease and market comparable were specified with stable identifiers and clear grain.

Address normalisation resolved properties. Tenant resolution used name normalisation, registered entity lookups and parent linkage, with a human review queue for uncertain matches.

The review queue mattered. Roughly one in eight tenant matches needed a person, and forcing those through automatically would have produced confident errors in exposure reporting.

Space was resolved last and was the fiddliest, because suites split and combine over time and the history has to survive that.

How were lease economics extracted from documents?

Through document processing that reads executed leases, amendments and side letters, extracting a defined field set rather than attempting general comprehension.

Extracted fields include commencement and expiry, base rent and escalation schedule, renewal and termination options with notice windows, recovery structure and improvement allowances.

Every extracted field carries a confidence value and a link to the page it came from. Low confidence fields route to a reviewer rather than entering the model silently.

Amendment chaining was the difficult part. The operative term is whatever the latest amendment says, and amendments are frequently filed separately from the original lease.

How does the renewal propensity model work?

It scores each active lease on the likelihood of renewal at expiry, refreshed monthly and surfaced from twelve months out.

Inputs include rent position against current market for comparable space, payment behaviour, service request history, occupancy density change, tenant sector conditions and remaining term.

Rent position against market turned out to be the strongest single factor, which surprised nobody, but its interaction with service history was not something the team had considered.

Each score is delivered with its main contributing factors, so an asset manager sees not just a ranking but the argument behind it.

How is capital prioritised?

By scoring each request against the income it protects, the rollover exposure at that asset, condition urgency and the effect on rent position.

That last factor connects capital spend to leasing outcomes, which had never been made explicit before.

The output is a ranked list with the scoring components visible, presented as an input to the quarterly meeting rather than as a decision.

Two requests moved sharply up the ranking in the first cycle, both at assets with concentrated rollover inside eighteen months.

Need decision support your team will argue with rather than ignore? Read more at AI consulting capability.

Technologies and Tools Used

The stack was sized for a lean team. Everything chosen had to be operable by two analysts after handover, which ruled out several options that would otherwise have fitted.

What did the data platform look like?

A cloud data warehouse holds the resolved entity model, loaded daily from the property management platform, the customer relationship system and the market feed.

Transformations are version controlled and tested. Every published figure traces back through the transformation chain to a source record.

The document store is processed on arrival rather than in batch, so a newly filed amendment updates lease economics within a day.

Data quality checks run before publication, and a failed check holds the refresh rather than releasing partial data into a dashboard.

How was document extraction built?

Layout aware document processing handles the scanned leases, with a defined extraction schema rather than open ended summarisation.

A language model performs field level interpretation where the wording varies, constrained to return values in a fixed schema with a confidence value.

Anything below the confidence threshold routes to a human review queue with the source page displayed alongside the proposed value.

Review decisions feed back as labelled examples, so extraction quality on the group's own lease templates improved across the engagement.

Which modelling and serving components were used?

Python for modelling, with interpretable gradient boosting and explanation output per record rather than global feature importance only.

Scores are recomputed monthly on a schedule and written to the warehouse, where the reporting layer picks them up.

Delivery to asset managers is through the business intelligence tool the group already used, which removed a training burden entirely.

A scenario screen allows a manager to test a rent assumption and see the effect on the propensity score before a negotiation.

Nothing bespoke was written where a maintained library would do, since two analysts cannot carry custom code they did not author.

What did the operating setup cover?

Scheduled refreshes, alerting on failed quality checks, and monitoring of extraction confidence rates by document template.

Model performance is tracked against actual renewal outcomes as expiries pass, which is the only honest measure available.

The human review queue has a service target, because extraction quality is worthless if reviewed items sit unactioned.

Documentation was written for the client's two analysts, and handover included working sessions on retraining and on adding a new source.

A short operations checklist was left behind covering the weekly refresh, what a healthy extraction rate looks like, and when to escalate.

Implementation Process

The engagement ran about eighteen weeks. Entity resolution took the first seven and was the phase that determined whether anything downstream would be trustworthy.

A deliberately small first scope, one asset type in one submarket, proved the approach before it was widened.

How was the first scope chosen?

Light industrial in a single submarket, chosen because leases are simpler, the asset count was meaningful and one asset manager owned all of it.

That last point mattered most. A single owner meant fast feedback and one person's judgement to test extraction results against.

Simpler leases also meant document extraction could be proven before it met the multi amendment office leases that would stress it.

The scope was widened to office and retail from week nine, once resolution and extraction had been validated on the easier set.

How was entity resolution validated?

By reconciling resolved occupancy and rent roll totals against the accounting platform, at asset level and at portfolio level.

Differences were investigated individually rather than tolerated. Three of the first eleven differences revealed genuine data errors in the source systems.

Tenant parent linkage was reviewed by the asset managers directly, since covenant exposure is the number they are held to.

Acceptance required a full reconciliation with every remaining difference explained, which took two attempts and was worth the delay.

Validation was repeated after the scope widened to office and retail, because multi tenant floors introduce space splits that industrial assets never surface.

How was extraction accuracy measured?

Against a manually abstracted sample of leases, field by field, with accuracy reported per field rather than as a single number.

Commencement and expiry dates extracted almost perfectly. Escalation schedules and recovery structures were considerably harder and needed schema refinement.

Option notice windows were the highest value field and among the harder ones, so those were routed to review regardless of confidence in the first release.

Reporting accuracy per field rather than overall kept the conversation honest and told the team exactly where to keep a person in the loop.

How was adoption handled with a lean team?

By delivering into the reporting tool the group already used, so nothing new had to be learned to see a score.

Asset managers were shown the contributing factors before they were shown the ranking, which framed the tool as an argument rather than an instruction.

The first monthly cycle was run jointly, with the delivery team present while the asset managers worked their own list.

One asset manager challenged three scores in that session. Two challenges were correct and led to a feature change, which did more for adoption than any demonstration.

Have four systems that disagree about the same tenant? Talk to us.

Business Results

The figures below are estimated business outcomes, modeled on the engagement scope and on published market benchmarks. They are projections rather than audited client results, and outcomes that could not be responsibly quantified are described in words.

Counts observed during the work are stated as observations rather than projections.

What changed about portfolio reporting?

Time to produce a quarterly portfolio review is projected to fall by an estimated 25 to 30 percent, mostly by removing manual lease reading and system reconciliation.

Tenant covenant exposure across all assets can now be produced on demand, where previously it required a person working across three systems.

Weighted average lease term and rollover exposure are computed rather than assembled, so they appear in every review rather than when somebody has time.

Every figure traces to a source record, which shortened the quarterly meeting considerably because the numbers stopped being disputed.

What changed about renewals?

Renewal conversations now begin around twelve months ahead on scored high risk leases, against roughly six months previously.

Renewal capture on assets where the score is acted on early is projected to improve by an estimated 10 to 15 percent.

Downtime, improvement allowances and leasing commissions associated with failed renewals are projected to reduce by an estimated 12 to 18 percent on those assets.

Two renewal options that had been missed in prior cycles were surfaced during extraction, both with notice windows inside the following year.

What changed about capital allocation?

Requests are ranked on a shared basis weighting income protected, rollover exposure, condition urgency and effect on rent position.

Two requests moved sharply up the first ranking, both at assets with concentrated rollover inside eighteen months. Neither would have been prioritised under the old process.

The group's suspicion that it was underspending on high rollover assets was confirmed by the first ranking rather than remaining a suspicion.

Capital planning discussion time is projected to reduce by an estimated 20 to 25 percent, since the ranking replaces the first half of the debate.

What changed structurally?

Lease economics now exist as structured, refreshed fields rather than as documents somebody has to read.

The entity model gives the group a single definition of tenant, space and lease that every system reconciles against.

Three genuine data errors in the source systems were found and corrected during reconciliation, which had been quietly distorting occupancy reporting.

Two analysts operate the platform without assistance, including adding a new market data source after handover.

Adding a new market data source after handover was done by the client team alone, which was the acceptance test everyone had agreed on at kickoff.

Want your quarterly review to take hours instead of days? Contact us.

Lessons Learned

The modelling was the easy part and the smallest part. Everything that mattered happened before a model was trained.

What surprised the client most?

That the same operating company appeared under four spellings and that two of its subsidiaries had been treated as unrelated tenants.

That reconciliation surfaced three real errors in systems everyone trusted. Those errors had been affecting occupancy reporting for some time.

That two renewal options with notice windows inside the following year had been sitting unnoticed in scanned amendments.

That rent position against market interacted with service request history in a way the team had never considered when judging renewal risk.

What worked better than expected?

Starting with one asset type in one submarket under one asset manager. Feedback was fast and there was one person to be accountable for judgement calls.

Reporting extraction accuracy per field rather than overall. It told the team precisely where a human still belonged.

Showing contributing factors before showing rankings. It changed how asset managers received the tool from the first session.

Delivering into the reporting tool the group already used, which removed the training burden that usually slows adoption in lean teams.

What would be done differently next time?

Amendment chaining would be designed at the start rather than discovered during office lease extraction. It is the hardest part of lease data and it arrived late.

The human review queue would have a service target from day one, since reviewed items sitting unactioned undermine the whole extraction argument.

Model performance measurement against actual renewal outcomes would be instrumented before launch rather than added afterwards.

Source system data quality would be reported back formally, since the errors found deserved a correction process rather than an informal fix.

What should other property groups check first?

Try to produce total exposure to a single tenant covenant across the whole portfolio. The time that takes is the size of the problem.

Check whether renewal and termination option notice windows exist as structured data anywhere. In most groups they do not.

Check whether occupancy calculated from the accounting platform matches occupancy calculated from the leasing system. Differences are usually informative.

Ask when renewal conversations typically begin. If the answer is six months, there is value sitting in the six months before that.

Frequently Asked Questions

What do data science consulting services deliver for a real estate portfolio?

Usually three things in sequence. A resolved data model that makes property, space, tenant and lease agree across systems. Lease economics extracted from documents into structured fields. Then decision products built on that foundation, such as renewal propensity scoring, rent position benchmarking and capital prioritisation. The first step consumes most of the effort.

Why is entity resolution such a large part of the work?

Because source systems identify the same things differently. Accounting platforms use internal codes, leasing systems use company names typed at deal stage, and the same operating company can appear under several spellings with subsidiaries that look unrelated. Until those records resolve to one identity, any portfolio aggregate is unreliable regardless of how good the modelling is.

Can lease terms really be extracted from scanned documents?

Yes, with a defined field schema rather than open ended summarisation, and with confidence values on every field. Commencement and expiry dates extract very reliably. Escalation schedules, recovery structures and option notice windows are harder. High value and lower confidence fields should route to a human reviewer with the source page shown alongside the proposed value.

What drives renewal propensity in a commercial portfolio?

Rent position against current market for comparable space is typically the strongest single factor. Payment behaviour, service request history, occupancy density change, tenant sector conditions and remaining term all contribute. Interactions matter too. In this engagement the interaction between rent position and service history was not something the team had previously considered.

Why does explainability matter more here than model accuracy?

Because asset managers will not act on a ranking they cannot argue with. A score that disagrees with experience and offers no reasons gets ignored permanently after the first occasion. Delivering the main contributing factors alongside each score, in language the team already uses, is what makes the output usable in a negotiation.

How far ahead should renewal risk be visible?

Twelve months is a workable target for commercial space, against the six months that is common practice. The difference matters because a failed renewal costs downtime, tenant improvement allowance and leasing commission. Six months rarely leaves enough room to change an outcome, whereas twelve months allows a genuine conversation to happen.

How is capital expenditure prioritised using portfolio data?

By scoring each request against the income it protects, rollover exposure at that asset, condition urgency and the effect on rent position. Connecting capital spend to leasing outcomes is the part most groups have never made explicit. The output should be presented as an input to the capital meeting, not as a decision that replaces it.

What size of team is needed to run this after handover?

This platform is operated by two analysts. That constraint shaped the technology choices throughout, and several otherwise suitable options were rejected because the group could not staff them. Delivering scores into the reporting tool already in use removed the training burden that usually defeats analytics adoption in lean teams.

Are the results in this case study real client figures?

The counts observed during the work, such as the missed renewal options and the source system errors found during reconciliation, are observations from the engagement. The percentage improvements are estimated business outcomes modeled on scope and published benchmarks, and every one is labelled as projected. Client identity is withheld.

How long does an engagement like this take?

This one ran about eighteen weeks, of which the first seven went on entity resolution. A deliberately narrow first scope, one asset type in one submarket under one asset manager, proved the approach before it was widened to office and retail. Timelines depend far more on data condition than on modelling complexity.

Conclusion

Data science consulting services changed how this real estate group makes decisions, but almost none of that came from a model. Seven of eighteen weeks went on making four systems agree about what a tenant is, what a space is and which lease governs which. Along the way the reconciliation found three genuine errors in systems everyone trusted and two renewal options with notice windows inside the following year that nobody had seen. Only after that did the modelling become worth doing, and even then explainability was treated as a harder requirement than accuracy, because asset managers do not act on rankings they cannot argue with. Projected outcomes include an estimated 25 to 30 percent reduction in the time to produce a portfolio review and an estimated 10 to 15 percent improvement in renewal capture where the score is acted on early. Those are modeled figures and they are labelled as such. What is not modeled is the structural change. Lease economics are now data, renewal risk is visible a year out, and two analysts run the whole thing. If producing total exposure to one tenant across your portfolio takes more than a few minutes today, that measurement is the business case.

Figures shown are estimated business outcomes modeled on the engagement scope and on published market benchmarks, not audited client data. Client identity and commercially sensitive details have been withheld.