We make predictions every day in one form or another. These predictions shape the decisions we make at all levels, from someone trying to get to work and guessing when the next bus will arrive, right the way through to a government forecasting economic growth and assessing how much (or how little) money it has to spend. 

Making predictions is difficult

While we make predictions all the time, making accurate predictions – ones that aren't just guesses but actually reflect what happens – is difficult. Especially when it comes to predicting something as complex as homelessness, in many cases experience and intuition are unlikely to be enough. While expert judgement can be invaluable in assessing whether someone is at risk of homelessness, experts are human. They have limited time and resources to assess the information in front of them, and they – like everyone – have biases and can make mistakes. 

In particular, expert judgement can only go so far when we’re trying to make more proactive predictions. Much of the focus on preventing homelessness in recent years has been on the 56-day statutory prevention window – where people are at imminent risk of homelessness and the local authority has to take reasonable steps to prevent this from happening. At this stage, experts can make a full assessment of the needs of someone in front of them.

But if we go further upstream, and try to prevent homelessness before people have even entered this prevention window, then the approach can shift. At this earlier stage, predicting who is at risk of homelessness is less likely to be a discussion with someone seeking help, and more likely to involve proactively scanning a range of unwieldy datasets. Looking at numbers on a screen and trying to assess an individual’s risk can feel like trying to find a needle in a haystack.

What are predictive models?

This is where predictive models come in. Predictive models use data (normally large amounts of it) to predict what will happen in the future. Often when we’re talking about these models, we’re talking about machine learning, where algorithms learn from data to identify patterns and relationships, and make predictions about the future. In areas ranging from healthcare to weather forecasting, these complex models can dramatically improve the accuracy of predictions – often surpassing traditional methods by identifying subtle patterns in the data that humans or simpler models might miss.

Let's consider what a model using this approach to predict the risk of experiencing homelessness might look like. To choose the data for our model, we’d look for factors that we think might be associated with homelessness. For example, things like council tax debt, rent arrears, interactions with various public services, and demographic information. We could then feed this information through a machine learning model, and after learning the patterns in the data the model would assign each person a score of how likely they are to experience homelessness. 

‍From prediction to action

While it’s useful and important to see how well these models can do at identifying who is at risk of homelessness, this isn’t where our interest in predictive models ends. What we’re really interested in, and where we can build on promising existing evidence, is how we can use predictions to proactively prevent homelessness. If we can predict who is at the greatest risk, then we can drive support further upstream, and get people timely and effective help to stop them from ever experiencing homelessness. 

Testing approaches that use predictive models 

We have good reason to believe that predictive models could be a real force for good in this way. Evidence from the United States suggests that predictive models can be used to successfully identify people at risk of homelessness. Closer to home, early evidence from the UK on using data to target a proactive approach shows signs of promise. The London Borough of Barking and Dagenham used data to identify residents behind on their council tax payments and offer them support. Residents they called were more likely to make a payment plan than similar residents who weren’t called. 

Maidstone Borough Council have also previously used predictive analytics, bringing together a variety of data sources to predict households at risk of homelessness. Households identified as at risk were offered proactive support, and the households that Maidstone contacted were less likely to experience homelessness. While these examples were not randomised controlled trials (often called the gold standard of evidence), they do provide positive signs, and open the door to the huge potential offered by predictive models paired with proactive engagement. 

Our Using Data to Prevent Homelessness project – one of the Centre for Homelessness Impact’s Test and Learn evaluations commissioned by the Ministry of Housing, Communities and Local Government –  seeks to build on this promising evidence. Using a randomised controlled trial, we’re looking to learn not just whether predictive models can accurately identify who is at risk of homelessness, but whether targeting support based on these models can reduce homelessness. To do this, we’re working with our evaluation partners, Verian Group and Simmetrica-Jacobs, and our delivery partner Xantura, who provide the predictive models and the OneView data science platform for combining council data. The project will operate in Barking and Dagenham, Stockport, Test Valley, and Newham, where the local authorities will proactively call households flagged as at-risk by the model to offer support.

And that's what we want to know more about. We don’t just want to know whether we can predict who will experience homelessness. We want to see if by targeting support to these people identified by a predictive model, we can prevent  homelessness. Results from this evaluation in 2026 will provide robust evidence on whether proactive support targeted by predictive models successfully reduced homelessness for at-risk households.

Balancing data quality and scalability 

This approach begs the question: What does it take to build a predictive model capable of accurately identifying people at risk of homelessness? This hinges not only on having data, but on having the right data - comprehensive, nuanced, and carefully curated. 

It’s hard to nail down exactly how much data is needed for this approach. For a rough indication, we can look at one of the most promising examples in this space, which comes from the United States. One study from Los Angeles looked at how predictive models can be used to identify people at risk of experiencing homelessness. In this case, the authors noted that their model needed roughly 50 different characteristics and indicators (or ‘features’, in the technical terminology) to get an acceptable performance, and 150 to 200 for optimal performance. ‘Features’ that appear to be useful for predicting homelessness include data on things like debt, benefits, and interactions with various public services. 

However, to look at the quantity of data alone would be misleading – it’s not enough to have a huge volume of data. It’s also important to have good-quality data and to manage it carefully. It’s critical to consider how securely the data is stored, who has access to it, and to address key privacy concerns. These considerations all form a necessary foundation for using data in predictive modelling. 

Moreover, predictive models are only as good as the quality of the data that’s fed into them, raising significant ethical and legal questions. For example, if models are trained on data that reflects historical biases against certain groups, then the models can learn, perpetuate, or even amplify these biases in their predictions. As the data science adage goes, “garbage in, garbage out”: the most cutting-edge model in the world will only be as good as the data you feed it, and it’s vital to ensure that these models are used ethically and responsibly, safeguarding individual privacy and mitigating bias. 

These considerations around quantity and quality of data have important implications when we’re thinking about the scalability of predictive models. We know there is huge variation in data maturity across local authorities. This means it’s important that the approach can be accessible to local authorities with less data to hand, otherwise it will have limited value. If the predictive approach can work even in cases where local authorities have less access to plentiful quality data, then this opens up a lot more doors for using predictive models more widely. For example, using simpler models that use a set of if-then rules, which can be developed with less reliance on large datasets, could be one way to bridge this gap.

There’s no straightforward answer here on the quantity and quality of data required to be able to take advantage of predictive models. But as part of our Test and Learn evaluation on this subject, we will learn more about what can be done with this predictive approach in cases where areas have lower levels of data maturity, or face barriers to accessing key data sources.

What’s next?

At the heart of our interest in these predictive models is offering support to people who need it. 

Evidence shows that people are (rightly) wary of algorithms in predictive models being used for sanctions, but have more encouraging attitudes towards using algorithms to support people. Using predictive models isn’t an exercise in seeing whether we can successfully predict something that’s difficult to predict. What we want is to use these algorithms to help us find people who need support, and act to give them the support that they need. As we increasingly look as a sector towards ways that we can prevent homelessness, predictive models may offer us a valuable tool to move efforts further upstream, targeting people who are at risk as early as we can, and in doing so ensure that they receive effective support to prevent them experiencing homelessness.

• Luke Arundel is Evidence and Data Lead at CHI