In late 2019, NHS England and NHS Improvement released an overview document of advanced forecasting techniques that can be used in demand and capacity modelling for NHS services.

As the document says, “Advanced forecasting techniques are commonly used in other industries but require specialist knowledge and training to use effectively”. And not just to use – even higher levels of expertise in statistics and programming are needed just to get to the starting line: to work out which technique should be selected for a particular task because it shows the greatest forecasting skill.

It isn’t always worth the effort. “In many cases”, the document explains, “a simple approach (e.g. ‘next year will be similar to last year’) will work well enough for planning purposes, and this is the approach used in the suite of tools developed by the National Demand and Capacity Programme for elective care.” Until now, Gooroo Planner has also leant on that approach, and for the same reason.

Covid has torn up the rule book

But since that document was published, covid has driven a coach and horses through those reassuring words. It may be years before we can say again with confidence that next year’s demand will be similar to last year’s.

Let us go further, and admit that the assumption was never entirely satisfactory. Demand at local level isn’t constant, nor does it rise steadily. Even in normal times, clinical practice has periods of relative stability, punctutated by occasional change. So does the way NHS data is collected. And so does the measured demand for healthcare.

Measuring what happened last year and using it as a basis for next year, even with adjustments at the margins, is not actually forecasting at all: it is projection. Projection has its place – the annual contracting round, for instance – and is useful if the goal is to model a set of assumptions and scenarios.

But if you want to know what is likely to happen in the real hospital, so that managers can take effective operational action to deal with foreseeable issues, then you need a proper forecasting technique that is responsive to recent changes in the level of demand as well as its regular cycles.

Getting it right post-covid

So a year ago, we teamed up with Professor Piotr Fryzlewicz, who is deputy head of the Department of Statistics at the London School of Economics and a specialist in time series analysis. We wanted to determine the most skilful forecasting techniques for NHS demand and capacity planning, so that we could integrate them into Gooroo Planner.

We decided to focus on forecasting non-elective demand, because the NHS has to respond to non-elective demand immediately, and cannot delay the activity via a waiting list as it does for elective care. We also looked at forecasting average length of stay, because non-electives are heavy users of beds, and beds are a prominent capacity bottleneck in most acute hospitals. And we wanted to evaluate each forecasting technique by trust and by specialty, because that is the most common level of detail used by NHS organisations in their Gooroo Planner modelling.

It took many months and a small fortune to extract the right data for the evaluation from NHS Digital, but eventually we got it: 2.5 million rows of statistics, containing both the non-elective demand and average length of stay, by trust, by specialty, by week, for the last 15 years.

If that sounds like a lot of data to analyse, it is. But we wanted to have full confidence in the conclusions when we went ahead and integrated the leading forecasting techniques into Gooroo Planner. So we had to do this properly, based on the most relevant and comprehensive data available, at the right level of detail.

Exhaustive evaluation

Professor Fryzlewicz identified a broad panel of suitable forecasting techniques, and conducted the highly specialised work of testing them all against the data. Modern data science programming languages make these techniques available in their statistical libraries, but you have to know what you are doing because it is easy to get statistical analysis wrong. (This is where it comes in really handy to be working with a Professor of Statistics.)

For those of you who are into the technicalities, he comprehensively evaluated autoregressive forecasting, seasonal autogregressive forecasting, classical exponential smoothing, and Holt-Winters smoothing (with and without trend and smoothing parameters). In each case, he also tested whether adaptive trend estimation and detrending would improve the skill of the forecasts. (Forecasting techniques go by various names, so if you have heard of ARMA/ARIMA then the above list includes two models in this class.).

Having determined the most skilful advanced forecasting techniques for the job, we then set about integrating them seamlessly into Gooroo Planner. UPDATE: We released this as a major upgrade to Gooroo Planner on 8th December 2020, together with the evaluation of the techniques considered (see the “Forecasting for NHS planning” paper here).

Responsive post-covid planning

What does this have to do with covid?

The pandemic has completely changed the patterns of demand for NHS care, and the virus will continue to affect the demand for services for some time to come. Nobody knows how demand will change over the coming months while the virus is still circulating, or what the eventual levels of demand will look like when everything settles down and (touch wood) we have a vaccine. In other words, last year is not a good guide to next year; a projection model can reflect our assumptions, but not much more.

On the other hand, the advanced forecasting techniques we are building into Gooroo Planner (which will be available in addition to the existing projection approach) will respond much more quickly to these changing patterns of demand as they shift to new levels. They can’t predict what the next shift in demand will look like – probably nobody can – but they will adapt well to what has happened recently. In these uncertain times, that responsiveness is going to be crucial for managing services.

Best of all, all the complicated stuff has been done for you, and you won’t need a degree in statistics to get the full benefit. We will, of course, let you know when it’s available.