Two decades of urban science have made the daily behaviour of a district measurable, and that measurement usefully complements, rather than replaces, the vacancy rates, take-up, rent and yield analysis on which office investment still rests.
The idea that a city can be read as data is older than the current wave of AI-generated interest in the subject. In 2006 Carlo Ratti and colleagues published "Mobile Landscapes", which mapped aggregated mobile phone usage across metropolitan Milan at different times of day, arguing that location data, if aggregated, "could become a powerful tool for urban analysis" [1]. The same year the MIT Senseable City Lab exhibited Real Time Rome at the Venice Architecture Biennale, described as "the first example of an urban-wide real-time monitoring system" combining telecommunications and transport records to understand patterns of daily life [2].
What followed was a methodological shift. In "Eigenplaces", Reades, Calabrese and Ratti connected Telecom Italia Mobile network data to a geography of commercial premises, using eigendecomposition to identify recurring patterns of mobile phone usage across the urban area [3]. Reviewing the lab's parallel campus work, Ratti notes the central finding: "every space on campus had a different usage pattern", a signature describing how it was used over the course of the day [4].
Two decades of that work were collected in Atlas of the Senseable City, written with Antoine Picon and published by Yale University Press in 2023, which charts "pollution, traffic, pedestrian flow, crowds, commuting patterns, and other elements of our daily urban experience" [5]. The claim is modest and useful: cities are legible in motion, not only in stock.
Institutional office underwriting rests on a small number of market variables: capital values, prime rents, take-up and net absorption, vacancy and development pipeline, walk-time to a main transport interchange, and a qualitative read on the building and its location. These are certainly valid variables to keep in consideration when valuing a commercial real estate investment. They are also mostly annual or quarterly, mostly building-level, and mostly silent on how a specific location has changed over time and how it is expected to develop in future.
There is a deeper limitation. Comparable analysis, whether on rent, yield or capital value, is by construction backward-looking and static. It tells you what someone was willing to pay for a broadly similar building in a broadly similar place, at a moment that has already passed, in a market that has already priced whatever was then known. It is a measure of consensus, not of trajectory. When a district is changing, which is precisely when the return is largest, the comparable set is drawn from the district as it was. Urban data is one of the few inputs available to an investor that describes the present continuously and can be extrapolated forward with some discipline.
Three additions are worth taking seriously.
First, rhythm. Urban analytics measures when a district is alive, not just how many people work there. Sulis, Manley, Zhong and Batty built a computational version of Jane Jacobs's diversity concept for London using Oyster smart card flows, decomposing it into "intensity, variability, and consistency, each measuring different temporal variations of mobility flows", validated against activity proxies [6]. Pintér and Felde, working with Budapest call detail records, derived a "wake-up time" metric and used it to estimate the operating hours of shopping centres, nightlife districts and workplaces [7]. A Senseable City Lab study of two Seoul districts combined street features with street-level mobile phone tracking to test vibrancy "with respect to age groups, time of day, and day types (weekends/weekdays)", finding relationships that vary by street typology [8].
For an investor this is not a curiosity. A district with a sharp weekday peak and an empty evening supports a different asset than one with an even rhythm: it changes which ground-floor uses are viable, what the food and beverage offer has to be, how much of the amenity provision is used often enough to justify its capital cost, and how a building performs on the two or three days that determine an occupier's perception of it. European office occupancy still stands at 61 per cent against a 70 per cent pre-pandemic benchmark, with mid-week days close to pre-pandemic levels [9]. A lease is signed on floor area, but the floor area is now used with a rhythm that traditional metrics do not capture, and a building designed for the average week will disappoint on the peak day.
Second, amenity as behaviour rather than inventory. Conventional analysis counts amenities. Mobility data shows whether people actually use them. In "Desirable streets", Salazar Miranda, Fan, Duarte and Ratti used thousands of pedestrian GPS trajectories in Boston to measure willingness to deviate from the shortest route, finding that desirable streets "have better access to public amenities such as parks, sidewalks, and urban furniture" and "more diverse business establishments" [10]. A later Senseable study of Stockholm, using call detail records for over one million users, found that areas with more libraries, educational facilities and restaurants host more income-diverse encounters, and that improved access to parks and services is associated with lower social segregation [11].
The investment consequence is a sharper brief for the design team. If the amenity that draws people is the street rather than the lobby, then capital spent on the interface between building and street, on the terrace, the colonnade, the publicly accessible ground floor, buys more occupier appeal than the same capital spent on an internal facility that is used twice a month. It also tells an investor which amenities to underwrite as rent-generating and which to treat as cost of entry, a distinction that agency research on London retrofits reaches independently.
Third, correction of assumptions. The 15-minute city is a good example of data disciplining a slogan. Using GPS traces from 40 million devices, Abbiasov and co-authors, including Ratti and Edward Glaeser, found that "the median US city resident makes only 12% of their daily trips" within a 15-minute walk of home, while access to local services explains "eighty percent of the variation in 15-minute usage across metropolitan areas" [12]. MIT's own summary notes that American urban dwellers travel seven to nine miles on average for commercial and recreational needs [13]. Proximity is not the same as local behaviour, and an amenity map is not a catchment.
For underwriting, this is a warning against a common shortcut. A schedule of what lies within a ten-minute walk of the front door is a description of supply, not of demand, and it routinely overstates the benefit an occupier will actually receive. Measured behaviour, not measured distance, is what should feed the assumption.
If a district's trajectory is what matters, the sharpest question an investor can ask is where the next concentration of high-value occupiers will appear. The economics here are unusually well established, and they operate at a scale much finer than a submarket.
Studying advertising agencies in Manhattan, Arzaghi and Henderson measured how the benefit of being near similar firms decays with distance. Moving outwards in 250-metre rings, "the coefficients are 0.020, 0.023, 0.0042, and then 0 beyond 750 metres" [14]. The authors describe the effective radius as walking distance: "interactions occur primarily within 500 metres, perhaps a 15-20 minute journey of elevator rides and walking" [14]. The effect appears even inside a single building: in a study of start-ups randomly assigned space in a co-working hub, firms "more than 20 meters apart on the same floor are indistinguishable from startups on different floors", with close proximity associated with a three percentage point higher probability of adopting a peer's technology [15]. Because that assignment was random, and because a natural experiment at the Jussieu campus in Paris found displaced laboratories became "3.5 times more likely to collaborate" once co-located [16], this is causal evidence rather than the usual self-selection.
Two qualifications. Wage-based studies find agglomeration effects operating out to about five miles, with an elasticity near 4.5 per cent that falls roughly fourfold beyond that distance [17], so the very short decay distances belong to information-intensive industries. And the crucial point for an owner: Arzaghi and Henderson conclude that the benefits of these spillovers "may be largely capitalized into rents rather than wages" [14]. The cluster premium accrues to the landlord.
King's Cross is the London case, and the documented sequence is narrower than the usual masterplan story. The trigger was a transport decision: the 1996 move of the Channel Tunnel Rail Link to St Pancras, which the developer calls "the catalyst for change" [18]. One developer was appointed in 2001 and the partnership became the single landowner across all 67 acres [18], and consent was granted in December 2006 after six years of negotiation, in a permission that "allowed 20 percent flexibility to vary the mix of uses within the total floor space" [19]. Single ownership plus a deliberately loose consent is what allowed the occupier mix to follow demand for the next fifteen years.
The order of arrivals matters too. Central Saint Martins was the first occupier in 2011 [20]. The Francis Crick Institute, a partnership of six research organisations, opened in 2016, and the Alan Turing Institute was created in 2015 inside the British Library [21][22]. Universal Music pre-let around 177,000 sq ft in 2015 [23], and Google's commitment of "more than one million square feet" came later still [24]. Institutions anchored the district before the corporates arrived, and a convening body, the Knowledge Quarter, formalised the cluster in 2014 around a one-mile radius [25].
Some of this is now observable in advance. A UK study using web-scraped firm data at postcode level, clustered at a 250-metre threshold, identifies 344 innovation hotspots, 115 of them in London, and isolates three predictors of whether a neighbourhood hosts one: rail accessibility to skilled labour, where roughly 40,000 additional reachable workers raise the odds by 27 per cent; proximity to a research-intensive university, worth 48 per cent within five kilometres; and proximity to a large high-technology employer, where each additional kilometre of separation cuts the probability by about 15 per cent [26]. The same study finds occupiers pay an 11 to 13 per cent office premium for hotspot locations, around £150 per square metre in London [26]. It also finds that these fine-grained clusters "are almost always melting pots of different sectors" [26], which argues against underwriting a district on a single-sector thesis.
A parallel literature supplies the logic for what comes next in a given place. The principle of relatedness holds that the probability a place enters a new activity is a function of the related activities already present, and that this probability "rises between eight-fold and twenty-fold when we move from an unrelated activity to a related activity" [27]. Read together, the practical shortlist of data an investor can use to anticipate cluster formation is short and concrete: rail isochrone counts of skilled workers, distance to research-intensive universities and to large technology employers, firm registrations clustered at street scale, relatedness density from patent and labour flows, business-listing churn, and mobile-phone visitation.
Karen Chapple's team at the University of Toronto School of Cities has tracked visits to 62 North American downtowns using mobile phone data, indexed to pre-pandemic levels and updated quarterly, extending to European cities in 2023 [28]. The associated study found downtowns with "high concentrations of professional services, information, and finance fields, high density, long commute times" recovered more slowly, while those weighted towards healthcare, education, arts and public administration did better [29]. Economic mix, measured through movement, predicted recovery better than lockdown length.
The implication for real estate analysis is the important part. A single, cheaply available behavioural dataset, indexed and updated quarterly, out-performed the variables the market was actually watching at the time. It identified the composition of a district's employment base as the operative risk factor, and it did so while the recovery was happening rather than in retrospect. That is the model for how urban data earns its place in an investment process: not as a score, but as an early, independent read on the variable that turns out to matter.
Pivo and Fisher examined more than 4,200 US office, apartment, retail and industrial assets from 2001 to 2008 using Walk Score, and found that on a 100-point scale a 10-point increase in walkability raised values by 1 to 9 per cent depending on property type, and that walkability "was associated with lower cap rates and higher incomes" [30].
Accessibility shows up in current market data too. Savills reports that offices "located just five minutes closer to major transport hubs command 6.7% higher rents on average globally" [9]. Cushman & Wakefield finds nearly 75 per cent of European office leasing over the past twelve months occurred in core CBDs, with core leasing volumes up 10 per cent year on year against a 4 per cent decline outside the core, and core vacancy at 7.1 per cent [31].
The most complete demonstration of what this looks like in practice comes from the MIT Senseable City Lab itself. In a 2020 study, Kang, Zhang, Peng, Gao, Rao, Duarte and Ratti asked how far the appreciation of house prices could be predicted from data about the city rather than data about the building [32].
The scale is worth stating. The study covers 21,928 houses in the Greater Boston area, with 125,000 house photographs and about 470,000 street view images. House prices and listing photographs came from the listings platform Redfin, covering appreciation between February 2014 and February 2019. To that the authors added five further layers: points of interest from SafeGraph within a 600-metre threshold, Google Street View images sampled every 100 metres along the road network with eight angles at each point, anonymised visitor patterns from a mobile panel covering "about 10% of total population with mobile devices in the United States", travel times between census block groups from Uber Movement, and census demographics. The imagery was not scored by hand or by category. A convolutional neural network was trained to sort images by price band, then used to extract a 512-dimensional feature vector per image, compressed by principal components, so the visual features are predictive but not directly interpretable.
The results make the argument, and they also bound it. A conventional linear hedonic model of the kind valuers have used for fifty years explained about 48 per cent of the variation in appreciation rates. A gradient boosting model on the enriched dataset explained about 74 per cent, with street view imagery the most valuable of the added layers [32]. Read the ablation honestly and most of that gain comes from moving to a non-linear model, with the new urban data reducing prediction error by a further increment of roughly 6 per cent. The lesson is not that photographs replace analysis. It is that a better-specified model plus city data beats the hedonic baseline consistently, and that the physical and behavioural character of the surroundings carries real information about future value.
The most persuasive evidence for an investor is a London study that tested exactly the thing an underwriter cares about: whether such a model works in a place it has never seen. Using 130,557 transactions, the authors trained on the rest of London and predicted Southwark, a borough held out entirely. Adding street view and aerial imagery raised out-of-sample explanatory power from 69 to 77 per cent and, in the authors' words, cut the accuracy lost when generalising to an unseen area "by two thirds" [33].
The same approach transfers to commercial property, though the analysis has to change, because the outcome is rent and yield rather than price appreciation, and because the data is thinner. The clearest example is again from the Senseable group. Applying a street-level greenness index derived from Google Street View to 1,414 Manhattan office transactions and 7,403 office leases between 2010 and 2017, the authors found "an 8.9% to 10.5% statistically, economically and positive transaction premium and a 5.6% to 7.8% rent premium" for offices on the greenest streets relative to the barest, controlling for submarket, building class, park and subway proximity [34]. That is a design and public realm variable, invisible in any rent table, priced by the market at close to a tenth of capital value.
Where machine learning has been applied to commercial transaction data instead of to the city, the gains are narrower but real, and the office is the hardest asset class. Boosting trees applied to 7,133 properties in the NCREIF index reduced the average error between appraised value and subsequent sale price from 11.12 per cent to 9.25 per cent, with location alone accounting for nearly 40 per cent of the explained error, and the improvement for offices specifically at 1.44 percentage points [35]. The models are effectively finding submarket boundaries that appraisers had not drawn.
What would make such a model trustworthy is now well understood, and it is worth being explicit, because this is where most proprietary efforts fail. Validation has to be spatial and temporal, not random: because neighbouring properties share almost all their urban features, conventional random cross-validation systematically overstates accuracy, and spatially aware resampling is "more dependable and can increase generalizability" [36]. The London held-out borough test is the standard to aim at. Data has to be contemporaneous with the outcome being explained rather than assembled from whatever vintage is available. And the model has to be understood as evidence rather than as a forecast: RICS, whose valuers ultimately sign the numbers, holds that automated models "will not be suitable for many types of valuation, including complex commercial valuations, until the appropriate data sources become available", chiefly because lease income has never been public in the way residential sale prices are [37].
That is a description of a frontier, not of a dead end. The residential work exists because the data exists. The commercial data is arriving now, through lease comparables platforms, mobility panels, imagery at national coverage and, increasingly, buildings that report their own occupancy. An investor who builds the discipline early, on evidence rather than on a score, will be reading the market with instruments that most of the market does not yet own.
Urban analytics is best understood as a second reading of the same asset. Capital value and rent comparable analysis, combined with take-up and vacancy analysis, describes how the market has already priced a location. Granular urban data describes how that location is actually used, at what hours, by whom, and whether the amenity offer around it is real or nominal, among many other things. Where the two readings agree, conviction rises cheaply. Where they disagree, the disagreement is the finding, and it usually points to a specific piece of diligence rather than a revised number.
For office repositioning in particular, the practical value is in matching the physical strategy to observed behaviour: a mid-week peak rather than a flat week, amenity that draws use rather than fills a floorplate, accessibility measured in minutes to an interchange rather than in adjectives, and proximity to the anchors that clusters actually form around, measured in hundreds of metres rather than in postcodes. None of this substitutes for conventional analysis by itself, but it powerfully complements it, and it lays the methodological foundations for future developments in commercial real estate valuation analysis.
This article is provided for information only and does not constitute investment advice or an offer or solicitation. Figures are drawn from the sources cited and were current at the time of writing.