The Unprotected Left (right) Turn

It is said the unprotected left turn (right turn in left-hand drive countries like Australia, but I will write as an American here) is hard for autonomous vehicles (AVs). (Even ignoring pedestrians, which magnify the complexity if there were to be treated as full-fledged users rather than an after-thought.)

With an unprotected left turn there is ambiguity about whether gaps between vehicles are large enough for the AV to squeeze through safely, and whether oncoming traffic will yield to an attempt to cross, particularly as the wait gets longer and longer and the passengers in the turning vehicle become more and more impatient.

It’s hard for humans too, left-turns comprise a quarter of all pedestrian crashes. So why do we have it? I.e. why don’t we protect the left turn.

It is  a matter of vehicle delay and storage space. On a one lane per direction road, or even a two lane per direction road, vehicles that are queued to turn could block vehicles that might otherwise go straight (or vice versa) when there is no turn bay and they don’t simultaneously have a green.  

We could have a phasing configuration which gave each approach (North, South, East, West (N, S, E, W)) its own green time.  In this  case, if flows were more or less equal between left turns and through/right movements, this might be the optimal solution. But if flows were dominated by one or the other, then it would be less than efficient. 

Alternatively, if we have turn bays (dedicated turn lanes) to keep vehicles out of each other’s way, we could have a configuration (N/S Through/Right, N/S Left, E/W Through/Right, E/W Left) which paired the turns. But turn bays use up lots of space that could be alternatively used for just about anything than temporarily storing cars.

And of course these could be mixed (N/S Through/Right, N/S Left, E, W) depending on relative flows.

But if the North flow > South or East flow > West (or vice versa), then these strategies will leave large gaps that could have been used by crossing traffic, but weren’t because the signal wasn’t timed for it.

With sufficient real-time information about flows, the signals could be adjusted to turn the lower flow approach to red when there are no vehicles approaching to protect the higher flow approach. This information requires knowing total approaches, but would be more accurate if the number on each turn (left, through, right) were also known, but this might be hard to discern simply by their location if the use of turn signals is imperfect, and there are too few dedicated lanes.

Update: We could prohibit left (right) turns. This is down in Moscow, so I understand. The left-turn ban at intersections is useful with low-rate flow turning left, assuming all left-turn vehicles are willing to do right turns several times to get to their destinations. But, this may impose a heavy burden of additional traffic on other road sections.

Or we could just have more roundabouts. These create other issues.

Porsche waiting to make a left turn, despite a presumably high value of time.

Sydney’s Traffic Returning to pre-COVID Levels as Driving Skills Deteriorate | SMH

Andrew Taylor in the SMH writes: Sydney’s traffic returning to pre-COVID levels as driving skills deteriorate

My quotes:

But David Levinson, professor of transport engineering at the University of Sydney, said: “In general low traffic levels lead to higher speeds and to more dangerous driving, so as traffic returns to normal, we may see lower rates of fatal and injury collisions per kilometre travelled.” …

Professor Levinson said empty CBD offices meant train carriages would also continue to be empty.

“Walking to work is also much more common for people in high-density areas, which are more likely still to be places where workers are effectively locked out,” he said. “And if the CBD is relatively empty, driving and parking there is not as difficult.”

My own sense is the trains are indeed pretty damn empty compared to the before-times, and driving still feels down (though not as down) in my observation. Traffic counts are taken at specific locations, and I am not sure any one data source is right on this.


How to value transport projects

Instead of measuring and monetising the fairy dust of `travel time savings’, a transport facility should be assessed on how much access it produces per unit of investmentAccess is the ease of reaching destinations. E.g. you might measure how many jobs (or restaurants or hospitals, etc.) can be reached in 30 minutes and/or $5 (or the dual of this measure, such as how many prospective patients an ambulance can reach in 12 minutes). A transport facility that increases access to destinations for a cost effectively is good. 

So the question is: does a streetcar or road or bike path enable people to reach more activities in less cost (time, money, aggravation, risk, negative externalities, etc.) than before, at a reasonable expenditure? (This cost includes the social and financial costs of building and providing the infrastructure). In short, are the upfront capital costsand ongoing maintenance and operations costs of the facility justified by the lower variable costs of its users? 

Sometimes (which is to say, often) transport projects are promoted for real estate. Real estate prices monetise the transport benefits (above what the user bears in time, money, and effort) in land value (time savings are not actually money, they become money through land value). We can build models that estimate the real estate value provided by additional accessibility.

So a better way of assessing the transport benefits is through real estate price uplift, as the market captures how people value the transport benefit. (We cannot simply add land prices to travel time and travel cost reductions, as that would be double counting). Places with higher access, and where access is more valuable, are more expensive and more productive and pay higher wages. We don’t really need to understand the detailed market mechanisms, nor attribute costs to detailed categories, the land market tells us how much access is worth, and transport models tell us how much access is created by a change to the network – from those two facts we can estimate the value created.

Because many projects are promoted by real estate interests, who presumably believe they will get the monetized benefits of those projects through higher land values, the public has a reasonable expectation that those interests pay for the costs of the project (that is, the tax incidence falls on the land owner). There are a variety of approaches, generally lumped as value sharing or value capture. The most general of these, a land value tax, originally promoted by Henry George, captures all of the uplift caused by all the access created by both transport investments and changes in the distribution of human activities.

From a project assessment point-of-view, land value uplift has often been part of the ‘wider economic benefits‘, which are optionally added after the value of travel time savings, which is considered the main benefit. ATAP for instance writes:

WEBs are improvements in economic welfare associated with changes in accessibility or land use that are not captured in traditional cost–benefit analysis (CBA). They arise from market imperfections, that is, prices of goods and services differing from costs to society as a whole. Reasons include economies of scale and scope, positive externalities, taxation and imperfect competition.

The international literature to date has concentrated on four types of WEBs that arise from major transport initiatives.

– WB1: Agglomeration economies — productivity gains from clustering by firms

– WB2: Labour market and tax impacts — productivity gains accruing to governments via the taxation system

– WB3: Output changes in imperfectly competitive markets — profit increases for firms

– WB4: Change in competition — gains to consumers and more efficient production.

ATAP goes on to write: 

“WEBs are only likely to be significant, and so worth estimating, for sizeable transport initiatives located in or improving access to large urban areas”

This logic is backwards. Because of induced demand, road projects rarely actually ‘save time’. Transit is often slower than car, so creating a project that induces someone from car to transit also doesn’t save time, but must nevertheless be preferred if people voluntarily switch.

Yet despite not ‘saving time’, these projects do create economic value. From a consumer perspective for instance, people can find a better fit for housing in the same travel effort (and may prefer to ride passively than to drive), or can engage in shopping activities that better match their desires in the same time window. From a producers perspective, WB1-WB4 from above are all embedded in land value. 

In reality, WEBs are the benefits of transport. If there were no productivity gains from clustering, we would not have cities and instead choose to be maximally spread out, and not need to be proximate in any sense. If there were no gains to consumers from competition, everyone would pay monopoly prices for everything, etc.

And these WEBs do not show up in ‘travel time savings’ but consistently show up in land value (CBDs are more expensive than suburbs are more expensive than rural areas). The WEBs are implicit in the land value uplift which occurs as a result the increased access. ‘Wider economic benefit’, properly measured as land value gains due to increased access, can and should be considered the primary benefit of new investment, not a speculative add-on aimed at juicing the numbers.

The consequence of properly and completely valuing benefits and full costs systematically may very well be a higher benefits estimate than a travel time savings-dominated metric would produce, which, if decision-making were rational, would justify more construction of public and active transport than would otherwise take place. A tax system that captured the land value that was thus created could relax whatever financing constraints currently limit that investment.



  • Wu, Hao, and Levinson, D. (2021) The Ensemble Approach to Forecasting: A Review and Synthesis. Transportation Research part C. Volume 132, 103357 [doi
      • Review and synthesize methods of ensemble forecasting with a unifying framework.
      • As decision support tools, ensemble models systematically account for uncertainties.
      • Ensemble methods can include combining models, data, and ensemble of ensembles.
      • Transport ensemble models have the potential for improving accuracy and reliability.
      ABSTRACT: Ensemble forecasting is a modeling approach that combines data sources, models of different types, with alternative assumptions, using distinct pattern recognition methods. The aim is to use all available information in predictions, without the limiting and arbitrary choices and dependencies resulting from a single statistical or machine learning approach or a single functional form, or results from a limited data source. Uncertainties are systematically accounted for. Outputs of ensemble models can be presented as a range of possibilities, to indicate the amount of uncertainty in modeling. We review methods and applications of ensemble models both within and outside of transport research. The review finds that ensemble forecasting generally improves forecast accuracy, robustness in many fields, particularly in weather forecasting where the method originated. We note that ensemble methods are highly siloed across different disciplines, and both the knowledge and application of ensemble forecasting are lacking in transport. In this paper we review and synthesize methods of ensemble forecasting with a unifying framework, categorizing ensemble methods into two broad and not mutually exclusive categories, namely combining models, and combining data; this framework further extends to ensembles of ensembles. We apply ensemble forecasting to transport related cases, which shows the potential of ensemble models in improving forecast accuracy and reliability. This paper sheds light on the apparatus of ensemble forecasting, which we hope contributes to the better understanding and wider adoption of ensemble models.
    • This paper is the first dissertation paper from Dr. Hao Wu’s Dissertation: Theory of Ensemble Forecasting – with Applications in Transport Modeling. Hao successfully defended last month. It’s hugely important for changing how modeling is done, instead of relying on the one best model, an ensemble of models is more accurate and more reliable. Transport modeling has spent decades developing advanced (and Nobel prize-winning) methods, but has fetishised a single model approach rather than embracing uncertainty and humility. This needs to change. [Hao is also, as far as I know, the first Transport Engineering PhD from the University of Sydney since JJC Bradfield, who designed the Harbour Bridge and the Sydney Trains network] “In 1924, Bradfield was awarded the degree of Doctor of Science (for a thesis titled “The city and suburban electric railways and the Sydney Harbour Bridge”, the first doctorate in engineering awarded by the University of Sydney.”
  • Allen, Jeff, Farber, Steven, Greaves, Stephen, Clifton, Geoffrey, Wu, Hao, Sarkar, Hao, and Levinson, D. (2021) Immigrant Settlement Patterns, Transit Accessibility, and Transit Use. Journal of Transport Geography. 96, 103187 [doi]
    • ABSTRACT: Public transit is immensely important among recent immigrants for enabling daily travel and activity participation. The objectives of this study are to examine whether immigrants settle in areas of high or low transit accessibility and how this affects transit mode share. This is analyzed via a novel comparison of two gateway cities: Sydney, Australia and Toronto, Canada. We find that in both cities, recent immigrants have greater levels of public transit accessibility to jobs, on average, than the overall population, but the geography of immigrant settlement is more suburbanized and less clustered around commuter rail in Toronto than in Sydney. Using logistic regression models with spatial filters, we find significant positive relationships between immigrant settlement patterns and transit mode share for commuting trips, after controlling for transit accessibility and other socio-economic factors, indicating an increased reliance on public transit by recent immigrants. Importantly, via a sensitivity analysis, we find that these effects are greatest in peripheral suburbs and rural areas, indicating that recent immigrants in these areas have more risks of transport-related social exclusion due to reliance on insufficient transit service.
  • El-Geneidy, Ahmed and Levinson, D. (2021) Making Accessibility Work in Practice Transport Reviews (online first) [doi]
    • ABSTRACT: Accessibility, the ease of reaching destination, is the most comprehensive land use and transport systems performance measure (Levinson & Wu, 2020; Wachs & Kumagai, 1973; Wu & Levinson, 2020). Accessibility has been applied in planning research since the 1950s (Hansen, 1959), and still today, we find major barriers to adopting it in practice (Handy, 2020). Advances in computing and software have enabled researchers to generate complex measures of accessibility with higher spatial and temporal resolutions moving accessibility research at a fast pace, while the implementation of accessibility, in practice, lags (Boisjoly & El-Geneidy, 2017). Even simple measures, such as the cumulative opportunities measures of accessibility, confront challenges in adoption.


Research by Others


  1. How long must someone be dead before we should stop referring to them as “the late so and so”? (reading newspaper article describing the “late Erik Erikson”, dead 27 years.) Or should we say the late Isaac Newton?
    • <1 year 18.6%
    • 1-4 years 34.3%
    • 5-9 years 11.4%
    • >10 years 35.7%
    The median is just under 5 years, so I will go with that. 

News & Opinion

Dr. Hao Wu

Congratulations to Hao Wu for “satisfying the requirements for the award of the degree of Doctor of Philosophy at the University of Sydney.”

Thesis Title: Theory of Ensemble Forecasting – with Applications in Transport Modelling 

Lead Supervisor: Professor David Levinson.


Ensemble forecasting is a modeling approach that internalizes uncertainties, combining models with different assumptions or pattern recognition methods, data from different sources, and different methods of combining models. Compared to the prevalent single-model procedure, ensemble model predictions are more useful as decision support tools.

The use of ensemble forecasting has significantly improved forecast accuracy in weather forecasting, and is increasingly adopted in other fields. We find a lack of awareness, or application of ensemble models in transport, so the benefits of ensemble forecasting are not being realized.

In this research we establish a systematic framework for ensemble forecasting, and propose the `ensemble of ensembles’ to combine uncertainties in different ensemble methods. Ensemble models are applied to transport-related cases to examine the performance of different ensemble methods, and to compare ensemble models with single-model forecasts.

We find ensemble models can improve forecast accuracy by a notable degree beyond the best single model. Simple and weighted average ensemble models have mixed results. Meta-learner ensemble models provide significant improvement upon base models, but require sufficient training data to calibrate. We find the linear meta-learner to be robust and have good performance even with small training data. Ensemble of ensembles method combining different ways of combining models improves performance upon ensemble models, and generally has the best performance.

We conclude that ensemble models, if properly applied, are able to improve model performance. We posit that transport modeling can benefit enormously from the wider adoption, and awareness of ensemble forecasting methods. We hope that this research opens the door to methodically adopting ensemble models into transport modeling, that future transport research can build upon.

Hao Wu

The first journal article published from the dissertation is:

  • Wu, Hao, and Levinson, D. (2021) The Ensemble Approach to Forecasting: A Review and Synthesis. Transportation Research part C. Volume 132, 103357 [doi]