News + Publications

This post was originally posted in two parts, which are combined below.

Part 1

Coanda recently had the opportunity to work with Calgary Homeless Foundation (CHF) to explore how mathematical modelling approaches could help better understand use of the various support programs and services for people experiencing homelessness.

CHF allocates resources for highest impact and outcomes for people at risk of or experiencing homelessness in Calgary. In partnership with all levels of government, agencies, and community leaders, they use data and research to advise on how best to leverage resources and programs in a unified fight against homelessness. Programs can range from street outreach to supportive housing. As part of funding these programs, CHF receives a variety of data for which they have created advanced data models. CHF has a vision of combining the program data with machine learning models to be able to forecast the usage and demand for the various programs and services. This would help enable better allocation of resources for the highest impact and outcomes, and in turn help streamline communication with funders.

Coanda sat down with CHF to learn about their data and data models, and to better understand some of the underlying “dynamics” that might exist in the various homelessness programs and services. Together we then brainstormed some modelling approaches for forecasting program usage or demand.

A few key challenges were encountered:

• Data for participant inflow into the ecosystem is limited
• Forecasting usage and demand are two very different things. Many programs are often operating at their maximum capacity, so demand for such a program is not necessarily reflected in the program usage numbers

After CHF and Coanda identified two main approaches that showed promise, Coanda developed a suggested approach to build progressively more sophisticated models. In recognition of CHF’s charitable status and the work they do, Coanda substantially subsidized the project. CHF is now in the process of training and testing these models. In our next post, we will elaborate on the models being considered, and how the challenges above are handled.

Part 2

In the previous post we introduced our work with Calgary Homeless Foundation to help develop data-driven forecasts for the usage and demand for programs and services in the Calgary Homeless Serving Sector of Care. Two main approaches were identified:

• A “macroscale” approach where we attempt to predict the usage or demand for a given program without looking at the details of who specifically is making use of the program. Here data is aggregated at a moment in time to provide population-level statistics and metrics that represent the state of the system at that point in time. Classical time-series forecasting methods (e.g., ARIMAX) or machine learning models, such as neural networks, can then be used to forecast usage or demand. With this strategy, the participant inflow would be implicitly captured through the time-series data, plus the inclusion of additional external variables.
• A “microscale” approach where forecasts are made on an individual level for each participant. These are then aggregated to get global usage or demand forecasts. This has the potential of generating more accurate and detailed forecasts; however, the data also needs to be “rich” enough for this to work well. In this case, a separate inflow model would be needed to predict the number and characteristics of participants entering the system. This inflow model could be similar to the macroscale models described above.

Finally, to make the distinction between forecasting usage, which factors in program capacity constraints, and forecasting demand, which is independent of capacity, we have to rely on the data being sufficiently expansive. If for a certain level of data granularity, the relevant conditionalities are not well-represented, then we can progressively reduce the granularity until we capture enough conditionalities. Then models should be able to learn the distinction between constrained usage and demand.

Putting Coanda’s data science expertise to use to understand how those experiencing homelessness engage with support services can hopefully help to enhance their effectiveness and reach.