Case study · Real estate investing

Industrial: which markets, and which buildings

I owned the models an industrial investment team used to screen markets and underwrite specific buildings: a market revenue ranking that says where to look, and a property leaseability model that says how fast a vacant building will lease. A research track tests new data sources that feed both.

Three tracks: market revenue ranking from rent momentum and a beta-regression occupancy forecast; property leaseability from a discrete-time survival model; and a research track testing alternative data.
Start with the market ranking to screen markets, then use the leaseability model to underwrite the specific building. The research track feeds new inputs back into both models.
Role
Model owner, from research to production
Scope
Top 100 US markets · national building universe · quarterly
Methods
Beta regression, discrete-time survival analysis, rank-correlation backtesting, geospatial features

1 Market revenue: which markets

For an industrial owner, revenue depends on two things: how fast rents grow and how much of the space is filled. This model ranks the top 100 US industrial markets by expected revenue three years out, combining both into one number, so the team can see at a glance where to focus.

Under the hood. The target is each market's realized revenue three years later, with rent growth and occupancy combined. Two signals are blended 50/50 and ranked across markets: a rent-growth momentum signal (trailing rent growth against the national average) and a forward occupancy forecast (current occupancy level and trend, plus a forward absorption estimate built from supply and construction pipeline data).

The occupancy forecast is a beta regression. Occupancy can never exceed 100%, and it varies less as a market nears full occupancy; beta regression respects both, while ordinary linear regression overshoots the ceiling and has to be clipped after the fact.

Scatter of market occupancy against a demand driver, with a straight linear fit that crosses 100 percent and a curved beta regression fit that approaches but never exceeds 100 percent.
Illustrative, synthetic data. The beta fit bends toward the ceiling; the linear fit runs through it.

Validation. Every quarter, the ranking is scored by its Spearman rank correlation with realized three-year revenue, across a backtest of more than 25 years. The blended ranking beat the previous rent-only model on ranking accuracy. It also holds up better through shocks, because momentum and occupancy tend to fail at different times.

Line chart of rolling rank correlation for rent momentum, occupancy forecast and their blend; during a market shock, momentum collapses while occupancy holds, so the blend dips far less.
Illustrative, synthetic data showing why blending two imperfect signals gives a steadier ranking.

2 Property leaseability: which buildings

Once a market looks attractive, the question becomes a specific building: if it is vacant, how quickly will it lease? The model scores every building in the national industrial universe each quarter with its probability of reaching stabilization (at least 90% leased) within eight quarters.

Under the hood. A discrete-time survival model estimates the chance a building leases in each quarter, given it is still vacant, and rolls those per-quarter hazards into a lease-up curve. Survival modeling fits the problem: lease-up depends on market conditions that change every quarter, and many buildings are still vacant, so their outcome is not yet known (censored). New construction (leasing from delivery) and re-lease (an existing building backfilling vacant space) are fit as separate models because their drivers differ. Accuracy is checked on discrimination (AUC) at several horizons. Earlier versions used gradient-boosted trees trained on synthetic sets from hedonic models.

Left: bars of the per-quarter lease hazard for new construction and re-lease. Right: the resulting cumulative lease-up curves, with an eight-quarter horizon marked.
Illustrative, synthetic data. The model predicts a hazard each quarter; the curve is what investors read.

From probability to decision. A probability on its own is hard to act on, so each building's score is also expressed as a percentile rank against three peer groups: all buildings nationally, buildings in the same market, and buildings of the same size class in the same market. Results can be cut by market, size and vintage class and tracked quarter by quarter.

Example output · made-up building

Shallow-bay warehouse · re-lease · built 1985

P(lease by 8Q)
74%
National pctl
81st
Market pctl
77th
Market × size pctl
70th

What drives both: geospatial features

Location matters in ways a building's own attributes do not capture. I built point-of-interest access features (highways, ports, airports) at census-tract level, plus 15-minute drive-time aggregates of nearby space fundamentals and lease-up events. A same-subtype variant keeps a warehouse's comparison set from being diluted by nearby flex buildings. These were among the largest single drivers of model lift.

Map sketch of a warehouse with an irregular 15-minute drive-time boundary stretched along a highway; same-subtype buildings inside are counted, other subtypes and buildings outside are excluded.
Illustrative. Drive time, not straight-line distance, defines the neighborhood, so it stretches along highways.

Research: does outside data help?

A framework to test whether outside data carries investment signal beyond standard fundamentals. Consumer card-panel spend, labor-cost data, maritime import shipments and freight flows were linked to individual properties through address matching and geographic-proxy joins, since none share an identifier with lease records.

Under the hood. Every source runs through the same engine: each period, split properties into terciles by their exposure to the signal, then track forward rent growth by group. A signal is useful only if the groups separate consistently. The strongest signals became ongoing inputs to investment decisions.

Bar chart of forward three-year rent growth by low, mid and high exposure to a signal, rising from low to high.
Illustrative, synthetic data showing the test, not a result.

Skills