Dosya

Long essays, thought through in questions and written with their sources.

Smart Cities, in Questions · 6/10

Can the twin foresee the city's future?

Updated · Published · Tansel Okay

In a period when overnight interest rates were high, we did an ATM cash-optimization study for a bank.

During that work, we found that cash withdrawals were affected (among many other factors) by the weather, too. Along the way, I learned that the weather matters for a great many economic activities.

The problem we were trying to solve was the kind where our chances of knowing any single person’s behavior were low. Still, we could predict a neighborhood’s behavior to the accuracy we wanted.

In the fourth essay I asked what the twin is; in the fifth, what feeds it. Now I come to the real claim, because one word in the twin’s definition carries weight: a predictive capability.1

What I have concluded so far is this: forecasting works when its subject hears the prediction and does not change its behavior; it fails when the subject does change. And what we talk about most in a city is exactly the kind that changes its behavior. If a forecast I make about traffic changes drivers’ behavior, could the result turn out better — or worse — than I expected?

Where it works

The weather forecast I relied on at the ATMs is the best example of forecasting that truly works.

On the city side, too, work falls into this class. The flood twin built for Galveston was validated by comparing its outputs with past flood events2 — a research prototype, but one of the few examples I can call a “validated forecast.” Istanbul’s disaster-focused twin is in the same class: an earthquake does not hear the prediction and change its plan.3 Dubai’s platform lists rainfall simulation among its use cases, but no result has been published; what we have is the institution’s own statement.4

Where it does not

Traffic is always spoken of as something that “simulates well.” What I saw in the second essay weakened that: on a circular track, a jam formed with no bottleneck at all. A bottleneck is only a trigger.5

But the real objection comes from elsewhere. Duranton and Turner find that, in US cities, as lane-kilometers rise, vehicle-kilometers rise right along with them; on the interstate highways in the densest parts of metropolitan areas, the increase is exactly proportional.6

Here the forecast is no longer a forecast. The sentence “this road will draw so many cars” becomes the justification for building the road; once the road is built, those cars come. The forecast changes its own subject.

Applying the same measure to the rest

I am not trying to count them all; this is an essay. But when I asked the same question of the others, I saw that I had already built the classes in the second essay. In 1948, Warren Weaver divided problems into three — problems of simplicity, disorganized complexity, organized complexity — and Jane Jacobs said the city is the third.7 The city’s forecasts fall into the same three.

Problems of simplicity: solar radiation, shade, wind, the spread of noise, a structure’s behavior under load. Few variables, a handful of equations. Helsinki’s model computes solar energy for nearly a million roof and wall surfaces;8 there are a million surfaces, but the calculation is always the same. Here the simulation runs quietly every day, and it would be unfair to belittle it.

Disorganized complexity: tomorrow’s electricity load, tomorrow’s water use. This turns out to be the name of what I had sensed at the ATMs. Weaver’s own example was an insurance company’s annual accounts; you do not need to know each molecule, the average is enough. And this class has now been measured: a meta-analysis that examined 421 forecast models from 59 studies finds that predicting a single household’s load is about 25 points more error-prone than predicting the whole system’s load.9 The individual is unforeseeable, the crowd foreseeable. The class Weaver drew in 1948 was put into numbers in 2023.

Organized complexity: traffic demand, population projection, a line’s ridership 20 years out, a project’s cost and benefit. In the second essay, I had already seen that traffic cannot be understood through averages. And this is exactly where forecasts fail systematically.

One thing does not fit into Weaver’s three boxes: weather, flood, earthquake. These are neither simple nor computed by averages; yet they sit on the side that works. What they share is something else: they pay no attention to me (to my forecasts).

So the measure is not in the thing itself; it is between me and that thing. I cannot change the weather; whatever I know, the rain will fall. I can change traffic, and I do — because I feed the forecast into a decision, and the decision pours concrete. The question is not “can it react”: do I have the power to change the thing; does knowing lead me to use that power?

Disorganized complexity is fragile for exactly this reason: the power over it may look unused, but it is in fact used. Electricity has a time-of-use tariff: different prices apply at different times of day. The regulation’s own rationale is written out plainly: so that the grids can be used efficiently.10 That is, the price is set to change the hour of consumption. The day, peak, and night lines on my bill record this.

So tomorrow’s load, too, is to some degree in the third class, organized complexity. The distributor forecasts the load, then shifts it with price; the next forecast becomes a forecast of a load shaped by its own price. Here, too, the loop closes on itself — only slower, and more polite.

So there are two separate questions. One: can this thing be foreseen? Its answer was given seventy years ago, in Weaver’s classes. The other: does foreseeing spoil the forecast? Weaver has no answer to this, because his question was about the structure of the problem; this question is between me and the problem. Its answer lies in power.

Thirty years

Flyvbjerg and colleagues examined 210 projects in 14 countries. In nine out of ten rail projects, the passenger forecast came out too high, averaging an overestimate of 106 percent. In half of the road projects, the gap between actual and forecast is greater than 20 percent. The truly striking thing: across the thirty years studied, the forecasts did not become more accurate.11

Thirty years. Computing, data, models — all increased. Accuracy did not.

There are two explanations; both must be named together. The first is psychological: the planner treats their own project as unique and special, and does not look at what happened to similar ones. This is called the “planning fallacy”; its source is the “inside view.”12 The second is incentive-driven: the numbers are bent on purpose so the project gets approved.13

To call the first a lie is unfair. To call the second an error is naïve.

Whom will the forecast convince?

This brought me to a question: what determines a forecast’s accuracy?

In Norway, forecasts for tolled and toll-free roads have been compared. The forecasts for tolled roads turned out fairly accurate; those for toll-free roads were inaccurate and generally too low, because they did not take into account that demand itself would grow. The explanation for the accuracy is striking: over the years, those planners had been scrutinized and pressed to give careful forecasts.14

On privately financed toll roads the direction reverses. There the party the forecast must convince is the financier; the optimism bias is pronounced.15

In all three there are the same models, the same software, and the same profession. The only thing that changes is who pays when it turns out wrong.

The remedy, and what the twin is doing there

The remedy Flyvbjerg proposes is not to model the project in more detail: it is to look at what happened to similar projects. Its name is reference class forecasting; its logic is the “outside view.”12

A digital twin is the inside view, perfected. It is built on increasing the detail of this city, this intersection, this building.

The twin and the outside view are not alternatives. The method does not throw out the inner detail; it puts it in order — first the base rate of similar projects, then this project’s detail. The problem is not the detail itself; it is where you start. A twin that starts from the detail becomes a very expensive wrapper for its own story.

Does a wrong forecast sometimes help?

The strongest objection stands at exactly this point.

In 1967, Albert Hirschman said this: the person who underestimates difficulties also underestimates their own creativity. The two errors cancel out; the project starts, a difficulty arises along the way, and somehow it gets solved. Better than never starting. He even gave it a name, derived from Adam Smith’s invisible hand: the Hiding Hand.16

A defense not of the accurate forecast, but of starting.

But the numbers say something else. Flyvbjerg tested the same claim against data from thousands of projects; in some three hundred where both cost and benefit were known, what Hirschman described really happens in one-fifth. In four-fifths the opposite occurs: on top of the cost overrun comes a benefit shortfall as well.17 There are objectors, too — one finds, across 161 World Bank projects, that the Hiding Hand wins.18 The debate today may be not about “does such a thing exist” but “is this the typical case.”

Hirschman draws attention to the fact that human creativity cannot be modeled. So giving the twin more detail does not raise the odds of landing in that one-fifth where the project is rescued along the way.

Who chooses the base rate?

The most uncomfortable question remains. The definition of “similar project” is elastic; it shifts according to whom the forecast must convince. I closed off three ways out — more detail, the outside view, the Hiding Hand — and each time came back here.

I have an answer from my own profession. In a solution-architecture document, two things are written down separately: all the known options, and why we chose the one we chose. That way someone who comes later can object on two distinct grounds — was the list incomplete (were all the possible solutions of that time considered?), were the weights of the rationale right (were the requirements captured, estimated, and evaluated correctly?). This does not make the project’s design choice objective. It makes it criticizable.

This is exactly what was missing in Hirschman’s eleven projects: he looked only at projects that went ahead — what statisticians call “sampling on the dependent variable” — so the rejected ones never entered the sample, and his selection could not be audited for fifty years.17

But I also saw this: the record’s job is not to capture the judgment; it is to limit what can be claimed in the judgment’s name. And a document works only if someone has a reason to open it.

So, can it foresee?

I can foresee a thing I have no power over; this should not be taken lightly: flood, earthquake, solar radiation, tomorrow’s electricity load. In a thing I have power over, and use that power on, I cannot separate the forecast from the intervention — but not because of a flaw in the twin.

A forecast’s accuracy does not lie in the simulator’s resolution; it lies in price and in the obligation to account. In Norway, the forecasts held because the planners were scrutinized. In the forecast that had to convince a financier, the optimism was pronounced. This does not mean the forecast is arbitrary: shared technical standards limit bias. But what they limit is not who is held to account.

A simulation’s output can look detailed and convincing on screen; but the error can emerge only after the work is done.

And this brings me to the next question: Who decides — planners, algorithms, or residents?

Correction (15 November 2026): In the first published version, Hirschman’s eleven projects were said to have had “no list of the rejected ones.” The critique in the source (Flyvbjerg) does not put it quite that way; the sentence was rewritten to match the critique’s own terms — “sampling on the dependent variable.” The argument is unchanged; the claim was brought closer to the source.

I wrote this essay with help from the Thinking Studio I built and from Claude, working from the sources.

Notes and sources

Each note gives the statement the text rests on, in the source’s own language; then the source.

note Statement (in the original language) Source
1 “A digital twin is a set of virtual information constructs that mimics the structure, context, and behavior of a natural, engineered, or social system (or system-of-systems), is dynamically updated with data from its physical twin, has a predictive capability, and informs decisions that realize value.” National Academies of Sciences, Engineering, and Medicine, Foundational Research Gaps and Future Directions for Digital Twins, 2024
2 “Validation of FlowsDT-Galveston is done by comparing its outputs based on historical flood events with reported flooded areas from social and news media during those events.” Mandal, Zou et al., “FlowsDT: A geospatial digital twin for navigating urban flood dynamics”, Computers, Environment and Urban Systems, 2026
3 “The pilot project will use digital twin technology to improve earthquake disaster management in Istanbul.” USTDA, 12 December 2024
4 “…practical applications in urban planning, infrastructure and asset management, rainfall simulation, and digital solutions that support decision-making” Government of Dubai Media Office, 2 July 2026 — the institution’s own statement; no result published
5 “Thus, a bottleneck is only a trigger and not the essential origin of a traffic jam.” · “We have performed an experiment on a circuit to show the emergence of a jam with no bottleneck.” Y. Sugiyama et al., “Traffic jams without bottlenecks”, New Journal of Physics 10, 2008, 033001
6 “We find that VKT increases proportionately to highways…” · “For interstate highways in the densest parts of metropolitan areas we find that vkt increases in exact proportion to highways, confirming the ‘fundamental law of highway congestion’ suggested by Downs (1962, 1992).” · “We conclude that an increased provision of roads or public transit is unlikely to relieve congestion.” Gilles Duranton, Matthew A. Turner, “The Fundamental Law of Road Congestion: Evidence from US Cities”, NBER WP 15376, 2009 · published version American Economic Review 101(6), 2011, 2616–2652. Data: US interstate highways, 1983–2003; “exact proportion” for the densest sections
7 “They are all problems which involve dealing simultaneously with a sizable number of factors which are interrelated into an organic whole. They are all, in the language here proposed, problems of organized complexity.” · problems of simplicity: “…problems which involved two variables or at most three or four…” · Jacobs: “Cities happen to be problems in organized complexity, like the life sciences.” · “The theorists of conventional modern city planning have consistently mistaken cities as problems of simplicity and of disorganized complexity…” W. Weaver, “Science and Complexity”, American Scientist 36(4), 1948, 536–544, p. 539 · J. Jacobs, The Death and Life of Great American Cities, Random House, 1961, ch. 22, pp. 433 and 435 (scan) — carried over from the second essay
8 “The model also includes solar energy calculations for all roof and wall surfaces in Helsinki … In total, there are nearly a million surfaces.” City of Helsinki, “Helsinki 3D”
9 “We use data from 421 forecast models published in 59 studies.” · “relative to system (reference category), our model estimates that the forecast errors on individual level are 25.16 percentage points higher.” · “the studies with individual-level prediction have a large variance and report significantly higher error values than the models with other grid levels.” · “By short-term time horizon, we mean load forecasts with up to one week ahead” Konstantin Hopf, Hannah Hartstang, Thorsten Staake, “Meta-Regression Analysis of Errors in Short-Term Electricity Load Forecasting”, Proceedings of the 14th ACM International Conference on Future Energy Systems (e-Energy ’23), 2023, DOI. The paper gives no absolute MAPE value; no percentage “accuracy” figure is stated in the text, only the individual–system difference
10 “b) Çok Zamanlı Tarife: Günün belirlenmiş olan farklı zaman dilimlerinde tüketilen elektrik enerjisi için, ilgili bileşenlerden oluşan ve farklı zaman dilimleri için farklı fiyat uygulanan tarifedir.” · rationale: “Bununla birlikte, iletim ve dağıtım şebekelerinin verimli kullanılabilmesi amacıyla sayaç altyapısının uygun olduğu tüketiciler, tüketici grupları, tüketim seviyeleri ya da bağlantı tipleri için farklı çok zamanlı tarifeler Kurul onayı ile uygulanabilir.” · time bands (on meters updated for summer time, year-round): “Gündüz dönemi: Saat 06-17” · “Puant dönemi : Saat 17-22” · “Gece dönemi : Saat 22-06” Energy Market Regulatory Authority (EPDK), Board Decision No. 5999-3, Dağıtım Lisansı Sahibi Tüzel Kişiler ve Görevli Tedarik Şirketlerinin Tarife Uygulamalarına İlişkin Usul ve Esaslar (Procedures and Principles on Tariff Applications), Article 20 — Official Gazette No. 29579 of 31/12/2015; EPDK’s latest published version. Winter bands (on meters not updated for summer time) differ: 07-18 / 18-23 / 23-07
11 “The sample used is the largest of its kind, covering 210 projects in 14 nations worth US$59 billion.” · “For nine out of ten rail projects passenger forecasts are overestimated; average overestimation is 106 percent.” · “For half of all road projects the difference between actual and forecasted traffic is more than plus/minus 20 percent.” · “Forecasts have not become more accurate over the 30-year period studied. If techniques and skills for arriving at accurate demand forecasts have improved over time, as often claimed by forecasters, this does not show in the data.” Bent Flyvbjerg, Mette K. Skamris Holm, Søren L. Buhl, “How (In)accurate Are Demand Forecasts in Public Works Projects?”, Journal of the American Planning Association 71(2), 2005, 131–146 — published abstract
12 “…managers make decisions based on delusional optimism rather than on a rational weighting of gains, losses, and probabilities.” · “These biases are often the result of the inside view in forecasting: decision makers have a strong tendency to consider problems as unique and thus focus on the particulars of the case at hand…” · “Adopting an outside view of the problem has been shown to mitigate delusion. It is applied by ignoring the specific details of the project at hand and uses a broad reference class of similar projects to forecast outcomes…” Bent Flyvbjerg, Massimo Garbuio, Dan Lovallo, “Delusion and Deception in Large Infrastructure Projects”, California Management Review 51(2), 2009, 170–193
13 “…decision-making is flawed by strategic misrepresentation or the presence of what economists refer to as principal-agent problems. Whereas the first explanation is psychological, the second is due to the different preferences and incentives of the actors in the system.” · “politicians, planners or project champions deliberately and strategically overestimate benefits and underestimate costs…” Same article
14 “…while traffic forecasts of tolled schemes are fairly accurate, traffic forecasts among toll free roads have a higher degree of inaccuracy and are generally underestimated.” · “An explanation for the observed discrepancy… is that road planners may have ignored the existence of induced traffic…” · “A likely explanation is that the planners over the years have been scrutinised and pressed to provide careful estimates for these projects.” Morten Welde, James Odeck, “Do Planners Get it Right? The Accuracy of Travel Demand Forecasting in Norway”, EJTIR 11(1), 2011, 80–95
15 Forecast error and optimism bias across more than 100 privately financed toll-road projects Robert Bain, “Error and optimism bias in toll road traffic forecasts”, Transportation 36(5), 2009, 469–482, DOI — citation only; full text not read, no figure cited. The authors explicitly write that it conflicts with Welde & Odeck (note 14); the direction varies with the tolling and financing model
16 “[S]ince we necessarily underestimate our creativity, it is desirable that we underestimate to a roughly similar extent the difficulties of the tasks we face so as to be tricked by these two offsetting underestimates into undertaking tasks that we can, but otherwise would not dare, tackle. … The Hiding Hand is essentially a way of inducing action through error, the error being an underestimate of the project’s costs or difficulties.” Albert O. Hirschman, Development Projects Observed (Brookings, 1967), pp. 13 and 29 — as relayed by Flyvbjerg; the book was not read
17 “statistical tests reject the principle of the Hiding Hand at an overwhelmingly high level of significance (p<0.0001)… the average project is in fact undermined by a double whammy of substantial cost overruns compounded by substantial benefit shortfalls.” · “the Malevolent Hiding Hand dominates by being present in 80 percent of projects.” · “Too-small sample. With a sample of just 11 World Bank projects spread over four continents, Hirschman’s dataset is much too limited to support the wide conclusions he drew.” · “…he committed the fallacy that statisticians call ‘sampling on the dependent variable.’” Bent Flyvbjerg, “The Fallacy of Beneficial Ignorance: A Test of Hirschman’s Hiding Hand”, World Development 84, 2016, 176–189 — accepted author manuscript. The “Malevolent Hiding Hand” name: Flyvbjerg & Sunstein, Social Research 83(4), 2016
18 “the author analyzes a sample of 161 World Bank-funded projects of different types and finds that the Hiding Hand prevails. While future research should ascertain this finding…” · Flyvbjerg’s reply: “I agree with my critics that the Hiding Hand exists… I show this happens in one fifth of projects… My main question was whether the Hiding Hand is ‘typical’… I show this is not the case, with 80 percent of projects not displaying Hiding Hand behavior.” Lavagnon A. Ika, World Development 103, 2018, 369–382 (DOI) · Bent Flyvbjerg, “Planning Fallacy or Hiding Hand”, World Development 103, 2018, 383–386 · also Graham Room, same issue, 366–368