An update from Grant Guesdon, MOVE 2.0 Lead
It's been 12 months since our last newsletter, and rest assured a lot has happened. At a broad level, we’re half-way through the 108 milestones that make up the development of MOVE 2.0.
We continue to collect data and build the algorithms that will power the platform. We’re starting to see how some of the individual parts work, which is exciting. But as with all IT projects as big as this one, there is still a lot to be done.
The first time the MOVE 2.0 subcommittee – the Delivery Group – will see the system (data and algorithms) will be July 2023. Once everyone has reviewed, we'll move on to the final stage: to update the system with the 2021 Australian Bureau of Statistics' census, and the latest mobility data that is available. What this will give us, by the end of November 2023, is the first complete set of hourly audience data for all signs nationwide for 365 days of the year.
We remain confident that the launch of MOVE 2.0 in early 2024 is on track. Over the last few months we have continued to consult with current MOVE users to find out what they would like to see and do with the data when we launch in 2024. We were pleased to learn that there are some requests that we can deliver early next year, and then others that we will aim to deliver at launch.
We are constrained to a large extent by the size of the data in MOVE 2.0: as you would expect, it is massive. The initial calculations suggest hourly audiences for 100,000 plus signs against a population synthesis of 2 million people; this is the equivalent of minute-by-minute ratings for 1,650 plus TV stations against a sample of 2 million people. That’s a lot of data at the one time! We need to ensure it’s in a reusable form, and that it doesn’t exceed the capabilities of the software for running reports.
While we have made a lot of headway, two issues remain unresolved: where we will source our mobile data, and writing a new methodology for place-based signs. It is hoped that both these matters will be resolved by the end of this year.
Mobile data investigation
On first glance, mobile sounds like a rich source of data. After all, it’s from a device people carry around and most Australians have one. But like all data inputs to MOVE 2.0, mobile data needs to be investigated and validated (checked for accuracy/inclusions and compared to other real-world data) to understand its part in measuring audiences.
To address this, we undertook an investigation into potential suppliers and the data they offered. After twelve months and a project cost of $400K, we now have a deeper understanding of the challenges of using mobile data. At a broad level, there are spatial accuracy limitations and demographic, geographic and temporal biases contained within the data. What this means is that using mobile data on its own does not deliver an accurate picture of movements at a micro level (ie. all individual locations).
Challenges with using mobile data
In terms of spatial accuracy, mobile data does not give us a precise location for individuals. This is because each mobile phone is trying to connect to the closest tower to make phone calls and access the internet. Therefore, what the mobile data is actually telling us is where individual phones are in terms of their proximity to the tower/towers. In the case of APPs which are plotting individuals via GPS they are only accurate within 5–20 metres of the location of the phone. You may have experienced this when using your ride share app, when your actual location in the real world doesn’t always neatly align with where the GPS marker says you are.
A 5–20 metre difference might not seem like much, but when it comes to audience measurement it’s enough to put you on the wrong side of the road which in turn changes how we score the visibility of our signs and more importantly it may incorrectly put you inside or outside the viewing area we give each sign in MOVE. Put more simply, a few metres could over or under report the volume of audience for the sign, which is significant for advertisers.
Spatial accuracy is only one consideration, in fact the main challenge with using mobile data is that it inherently contains demographic, geographic and temporal biases which affect its usability in an industry wide audience measurement system (refer image one).
For example:
About 90 per cent of Australians have a mobile phone, but ownership is under indexed among people aged 65+ and children aged 5 to 12. Meaning APP/SDK based mobile data has a bias when compared to the population. A problem that can’t simply be solved by going with a mobile services provider (over an APP aggregator or SDK solution) as they too have their own skews on the demographics for which they provide data. What we are looking for always in MOVE is a sample of data that can be validated against the Australian population.
Image one: Demographic, Geographic and Temporal challenges with mobile data.
But even if we could overcome demographic biases in the data (not easily done when the data is anonymised), it remains that:
- it’s at best only a two-dimensional picture (ie. can’t distinguish between floors at a location),
- not everyone takes their mobile phone with them everywhere they go (according to a Telstra survey),
- data dropouts and network outages will occur as seen throughout our investigation (eg. a train station test over a year revealed one month when half the average number of people were present due to a network outage in the area).
Our conclusion was that mobile data shows an inaccurate picture of total travel movements of individual people. However, this doesn’t mean it cannot play a role: we will use it as one of many data set to measure all Out of Home formats nationwide with accurate seasonal audience at an hourly level. This can only be achieved by incorporating a wide variety of data sets to model how audiences move around the country. This model is built based on real world behaviour which is then calibrated to relevant real-world data (where it is available).
How mobile data will be used to benefit MOVE 2.0 at a broad level
- Temporal profiles such as time of day, day of week for different types of locations (by grouping a representative sample of locations together),could be applied to audience volumes from other sources to distribute audience by hour across the week.
- Broad area people movements (eg. number of unique people present in the data in a particular area, after factoring down for device ID changes such as new phones, migration etc) will assist with calibrating the model we are using.
This is currently being prototyped in MOVE 2.0 for review by the Delivery Group in Q2 2023 and is expected to be included in audience data delivered in November 2023.
We also expect to finalise our investigation into mobile data suppliers and select the data source we will go with at launch by the end of 2022.
Development progress
The mobile investigation and prototyping are just a couple of the many milestones we have tackled in the last twelve months. Some of the other major milestones are outlined in more detail below, as well as listed in the system architecture summary diagram (refer image three).
Image three: Summary Architecture Summary, with 2022 development progress.
Synthetic population
MOVE 2.0 will use a total synthetic population of 2.5 million individual Australians of which approximately 2 million are aged 14+ years, which is the audience on which MOVE 2.0 will report. Each person in the synthetic population aged over 14 years will make their own individual trips (relevant to who they are) across 365 days of the year.
This year we’ve finished development of the first synthetic population based on the 2016 census and reweighted it in separate steps to both the Australian Bureau of Statistics (ABS) releases of estimated residential populations (ERP) and OZTAM universe estimates to ensure it’s the most accurate picture it can be during development. The 2021 census will be added prior to launch and follow the same re-weighting steps.
Image four: Demographic attributes of each synthetic population member aged 14 years an over.
Behaviour data
To move the synthetic population around we need real world information to inform how this population moves. This includes surveys that collect peoples’ movements, count data and mobile data.
The survey collection started in January 2022 following lockdowns lifting in NSW and Victoria with deployment of the Multi Sensor Tracking (MST) survey, which is carried by a person for 14 days. This survey provides the most detailed picture of people movements possible. This includes collection of demographic and usual behaviour information via questionnaire, then the MST tracks the person second by second indoors and outdoors over 14 days. In addition, the people carrying the device give us detailed information on why they are at each location using a phone App (eg. Person is at the shopping centre to do some personal business ie. banking).
The MST device collects GPS location (longitude/latitude) as well as second by second the acceleration, direction, elevation, temperature, and pressure. Therefore, what is tracked is the routes people are taking, seamlessly indoors and outdoors, which is not possible with GPS alone.
We’ve also built a database that is updated on a regular basis with all the latest traffic volumes, pedestrian and cyclist volumes, public transport routes and stops, airport passenger counts, and available shopping centre visitation counts.
We have available 36,000 location counts covering 23 million data points across all time periods across a year.
People movement
Obviously, this is a major part of our modelling and also the most complex. The model is based in two parts an Activity Based Model (ABM) and an Out of Market Model (OMM).
The ABM models all trips of the synthetic population when they’re within their local market for work, shopping, socialising etc. including any trips that may occur on the day an OMM trip commences or finishes to create a seamless picture of their travel.
For example, if a person in Sydney is travelling at 4pm to the airport to catch a 6pm flight to Adelaide from Sydney, and prior to that they’ve doing some shopping at various locations in Sydney earlier in the day. The trip to the airport and flight to Adelaide is covered by the Out of Market model, and trips in Sydney prior to this by the ABM.
The Out of Market model determines when people make a day trip of more than 150km from their home SA1 or an overnight trip of more than 40km (as a shorter distance as this aligns with the Federal Government National Visitor Survey definition of an overnight trip), and models that trip based on the purpose (business, holidays or visiting family/friends), mode (air, car or other) based on their demographics ie. age, income etc).
Inputs for models we've built this year
- A transport network that includes airports, every road (7 million road links totalling 3.7 million kilometres) and public transport options (86,500 public transport stops/stations covering 17,700 types of route service).
- Two national zoning systems. The first based on SA1 areas (used in the ABM) containing over 62,000 separate zones and the second based on SA2s (for the OMM) that contains 2,500 separate zones. Within the SAI areas we have also split CBD areas to smaller mesh block zones and separate zones for each shopping centre and airport. The movement models will MOVE people between and within zones to align with real world behaviour.
- Lastly, we’ve built a route inference model that produces a series of the plausible routes between zones based on the time of day, day of week and mode choice available.
Using the above inputs and the behaviour data we’ve built the initial versions of the ABM and OMM that will cover people movements.
As we’re about halfway through collecting and processing the survey data we’re starting to see these models work. When complete they will provide a detailed and comprehensive picture of how Australians travel.
As part of finishing these models, still to come is the incorporation of mobility data from mobile working alongside the survey data collected and volume controls using the real-world count data.
What’s next
Once the models are complete and validated against real world audiences, we will then have the ability to produce audience movements.
Internally we are working on site classification data, which is how each sign is classified in the system, from size to type to location. Signs will be classified as Roadside, Airport, Station, and Retail environments for an interim output in March 2023. Then, at the end of July 2023, we will see all formats (Transport and Place-Based) in the full pipeline of algorithms with the data working at once.
As you can see, if you have read this far, MOVE 2.0 is a herculean project but also an invigorating one. There are only thirteen months to go before the 2023 November deadline for the platform to be built, including updating to the latest available data inputs, sign details, and finalising audience data for launch in 2024.
Grant Guesdon, MOVE 2.0 Lead, Outdoor Media Association
