Where are the low-hanging fruits?
Mapping potential consumers of shared e-mobility to help decarbonise cities
In marketing research, the concept of ‘low-hanging fruits’ refers to potential consumers who are easiest to seduce. Focusing efforts on this group can maximise the success of marketing campaigns and profitability. In mobility planning, this concept could [and should] be adopted more often to achieve sustainability goals.
Imagine that a start-up just launched this new model of shared e-scooters in a busy town like Rotterdam. It’s natural to expect that, for the sake of financial sustainability, a significant part of the revenue should come from neighbourhoods composed by potential users and other factors of success [e.g. commercial activities, jobs, infrastructure][see figure below].
However, if shared e-mobility is meant to cause significant and positive impact on sustainability, helping cities achieve their goals, further structural changes in travel habits are certainly necessary. In short, part of the ‘unnecessary’ car trips should be more often replaced by more sustainable modes, like shared e-mobility. ‘Unnecessary’ is regarded here as a car trip that has similar profile [e.g. length, travel time, socioeconomics] of a shared e-mobility trip, and therefore could be ‘avoided’ or ‘replaced’. The individuals making those trips are the ‘low-hanging fruits’.They have a similar ‘trip profile’ of shared mobility customers, but are ‘not yet consuming the product’.
How to find low-hanging fruits?
In this article, an approach is proposed to help providers and cities find them. The approach is operationalised in the context of the Netherlands, a country where shared e-mobility is spreading quickly. Some use cases of this exercise can be applied to: (i) size potential market for expansions [e.g. deployment of vehicles or installation of facilities]; (ii) size potential impacts of modal shift on city-wide Co2 emissions; (iii) design subsidies that encourage providers to deploy assets in certain areas; (iv) change fees depending on the potential to attract former private vehicle users; (v) investigate reasons behind the existence of avoidable car trips; among others [the more, the better!].The steps are operationalised as follows:
- Defining how an ‘avoidable’ car trip looks like in the context of a given population [city, region, country]. This can be done, for instance, by looking at how users of shared e-mobility travel [e.g. trip distance, duration] and what are their characteristics [e.g. age, gender, income].
- Mapping where the avoidable car trips are produced. Many countries keep their Household Travel Surveys up to date so that city planners can use that information to have a ‘region-wide idea’ of travel habits [desire lines, purpose, mode choice, etc];
- Labelling locations in regard to their likelihood of having more or less low-hanging fruits. In this step, I apply Unsupervised Learning (k-means) to find probable clusters of low-hanging fruits. It’s a combination of avoidable car trips with relevant census variables.
In order to achieve (1), we used an anonymous, ‘privately acquired’ shared mobility OD travel matrix produced in 2020 by a third party mobility company. This OD refers to trips done by e-scooter users of Rotterdam during the summer of 2020 (June, July and August), and looks like this:
We can also visualise the OD as desire lines like these:
For (2), we explored the latest Dutch Household Travel Survey (2020) and combined it with (1). This kind of survey provides annual information about daily travel patterns of thousands of people [OD, mode, purpose, distance, duration…]. Last year for example , data about 62 940 individuals across the Netherlands were collected. The Dutch HTS can also be expanded to mitigate negative impacts of data collection biases and be a reasonable representation of how the whole population chooses to travel in a daily basis. An OD like the one below can be generated from the HTS:
In (3), we combine insights extracted from the HTS (2) with Census data to perform the unsupervised classification of locations. For the sake of simplicity, only a few socioeconomic variables were explored.
1. Profile of shared e-mobility trips
To define the profile of shared e-mobility consumers, we explored how far users travel. To do that, the acquired OD was transformed in geographic desire lines, from which simplified euclidean distances were calculated. In the histogram below, it’s possible to see that the majority of trips made on e-scooters (around 80%) are between 500m and 2,5km. A better approximation of the histogram to reality would be a curve shifted a bit to the right, due to network effects on distance, however this information was not available.
If compared to the distance of trip legs [smaller parts of a full journey] reported in the Dutch HTS, we can see that our target trip profile corresponds to at most 20% trip legs made by private vehicles.
Reports about the profile of shared mobility users often point to individuals of higher income and higher education levels as the likely users of shared services, which of course can spark discussions about the role of providers in promoting inclusivity… but for the sake of simplicity, we target individuals aged between 24 and 44 years old.
2. Mapping clusters of ‘avoidable’ car trips
Now that we know what is our target profile, let’s filter from the HTS database only journeys that are composed by at least one leg between 500m and 2,5km made by car. To those filtered journeys we apply expansion factors to (i) correct for collection bias and (ii) to expand the survey sample to overall population figures, and have an idea of ‘how big is the market’ for such a profile. In the figures below, we plotted a snapshot of the regions in the Netherlands that, in a typical day of 2020, produced our target trips. These were not categorised by purpose for simplification. The colours and sizes of the circles [figure below] indicate the number of target trips that are produced per Zip Code in a daily basis.
Looking at the spatial distribution of the target trips, we can realise that they cluster either a bit far away (outside a 5km radius) from city centres or at smaller cities/ villages in the metropolitan regions of larger cities (e.g. southeast and east of Rotterdam Centrum, west and south of Amsterdam in direction to Haarlem and Amstelveen respectively). Proximity to services, retail and amenities can explain part of those differences, since dense and diverse neighbourhoods tend to encourage shorter trips, which can be made by bicycle or by foot.
3. Finally, where are the ‘low-hanging fruits’?
After identifying the zip codes that produce our target trips, we added census information to the data set (density, age and proximity to amenities) and applied a K-means algorithm to allocate zip codes to different groups.
Conceptually, clusters that generate high number of target trips, densely populated and accessible to amenities (shops, markets…) can be considered our low-hanging fruits. To use k-means clustering, it’s necessary to first indicate the optimal number of clusters (k) to be generated. Among the available approaches, the Elbow method was chosen. Based on the result below, k = 4 is a reasonable number to group zip codes across the selected variables. After several iterations, the k-means algorithm was able to assign the locations to their respective groups. Now it’s time to label the clusters!
After examining the centres of each computed cluster, the Low-hanging fruits [high producers of avoidable car trips/ dense neighbourhoods/ a lot of young inhabitants] were identified as cluster number 1, and correspond to approximately 1100 zip codes (25% of the official 4000).
This initial clustering provides a reasonable universe of locations to apply more specific filtering techniques and to map potential consumers of shared mobility more accurately.
Let’s say that a mobility provider wants to attract ‘low-hanging fruits’, but still keep the supply of vehicles in areas that are ‘sufficiently urban’. Or city planners want to apply certain policies to incentivise use of shared mobility within municipal limits. In such cases, others spatial filters can be applied to the identified clusters [e.g. establishing a maximum distance from given zip codes to city centre, focusing on locations within given administrative limits, establishing minimum density levels].
We propose and operationalise an approach to help cities and mobility providers identify potential users of shared mobility probably willing to replace private vehicles with more sustainable options due to their trip profile and socioeconomic characteristics. If shared mobility could seduce more low-hanging fruits, more significant environmental impacts from modal shift could be achieved. This is not intended to look like an ‘ultimate solution’, but just contribute to the discussion of how to better insert shared e-mobility in our cities and cause significant impact.