Moving up the modelling value chain

Over the past few months I have seen repeated references to the image of the Gartner Analytic Value Escalator, or the Gartner Analytic Continuum. I haven’t found an original reference, and the image is sometimes drawn slightly differently, but the message is always the same: adding value to data, moving from hindsight via insight to foresight, by progressing from Descriptive to Diagnostic to Predictive to Prescriptive Analytics. There’s a good version of the image here

Isn’t this what we do in transport modelling? Well, yes and no. For example, often our data collection and analysis doesn’t progress much beyond Descriptive Analytics, and our forecasts are Predictive and rarely Prescriptive. Hampered by the skills shortage, constrained by WebTAG, limited by time and money budgets? Maybe. But I sense a change, both in what end users are looking for, and what modellers themselves want to do. If you haven’t read the recent Commission on Travel Demand report “All Change”, I suggest you do so ASAP. It makes answering the question ‘what will happen’ that more difficult and perhaps less relevant.  It opens up the option of exploring the question ‘what do we want to happen’ and ‘how do we make it happen’ a more reasonable proposition.

Transport modellers have traditionally not played a strong role in this sphere. Remember the “The model made me do it” cartoon (apologies – I don’t have an original reference for this either)? But we should and we can. Models don’t need to be more sophisticated to explain the impacts of different assumptions about the future, and explore the kinds of future that might result; and perhaps even simpler analyses will provide even more insights. Here are four steps that might help us get there:

• Collect and analyse the best data. Despite the increasing availability of Big Data, much of the model data that we collect and use is still through traditional flow and speed surveys, and if we collect more advanced data, it tends to get analysed at an aggregate level, usually as an input to replace traditional manually collected surveys. Yes, we use cameras rather than tubes; yes we use apps to collate household data. But still a lot of the detail, of the nuance is lost, basically because we aren’t looking for more and stick to the Descriptive stage. On top of this, the profession remains wedded to cross-sectional analysis, which makes it difficult to detect and understand change; and isn’t understanding change our biggest current challenge?
• Reinstate a much stronger diagnostics approach to data analytics. We talk about future mobility, we worry about the impact of autonomous vehicles on our forecasts, or about how to represent Mobility as a Service in our models. Much is indeed unknown, but some of these things are being trialled, and signals are emerging. Not enough learning (for example at Universities) makes its way into mainstream travel demand forecasting. Of course there is always the danger of observing correlation rather than causation, but from diagnosing correlation we may be able to start understanding the underlying causes. Too much conjecture, not enough robust statistical analysis.
• Explore up-front the desirable and undesirable futures. This is where we hopefully can and already have moved away from the predict and provide paradigm. I doubt any modeller will feel comfortable with a responsibility to predict, nowadays, allowing for just high and low growth assumptions. What are the plausible futures? How can we model them? Model results, properly acknowledging assumptions and limitations, allow us to engage in the debate and contribute with quantitative evidence on the circumstances under which these futures may occur.
• Use models to identify projects and policies that help or hinder getting there; dare move into the Prescriptive domain. The debate on autonomous vehicles is a useful example. Although we don’t yet know how people will value their time spent in these, we can calculate with existing knowledge and tools what would happen under different assumptions to the demand for travel (by car and overall, in a continuum from a value of travel time of zero to the current values). We can use current models how the extent to which travellers will be willing to share rather than own (from 100% to 0%) might affect travel demand, and the additional mileage that will be travelled by empty vehicles (I read recently that the average Uber carries 0.6 passengers (again without a reference, sorry!)). We don’t need highly advanced and new models to test this – many of the mechanisms I describe are covered by existing tools that exist in most major cities.

Alternative models representing quite different mechanisms, also may show quite different emerging futures. The 2017 RethinkX report “Rethinking Transportation” uses system dynamics and a number of self-reinforcing loops to end up in a highly autonomous and shared future, and quickly; others expect inertia and damping rebound effects to slow down progress and uptake of new technology. Both can be and have been modelled. In times of uncertainty, it isn’t the case that there is no value in modelling, in demand forecasting. Quite the opposite, it is the sensitivity testing of alternative input assumptions, of different world views, indeed of different model approaches, that moves modelling up the value chain and will start to throw light on what appears to be anybody’s guess at the moment.