How about hyperlocal pandemic forecasting

19.39, Tuesday 12 May 2020

I’m a big fan of weather forecasts. It’s an incredible feat to describe the ever-changing multi-variable fluid-dynamical state of the freaking atmosphere in such simple terms that we can

  • act – imagine how insanely complex it must be just to say “it’s going to rain today,” yet the entire forecast can be summarised in that line and I know whether to wear a coat.
  • communicate – I can hear or read a weather forecast and share it with another person, using speech only, face to face or just over the phone. I can’t even figure out how to share a URL to a Facebook post half the time.
  • reason – seeing a warm front on a map lets me predict for myself the next 24 hours, despite me having no idea about the actual numbers in the underlying atmospherics equations.


I’ve just been looking at some stats. Next day forecasts are 80% accurate, up from 66% a decade ago. The UK Met Office’s 4 day forecast today is as accurate as the 1 day in 1980.

Barometers: especially good. With their dial running stormy/rain/change/fair/etc. See some antique ones here. It’s everything you need to know in a single instrument: e.g. it was originally raining and it is improving quickly. A vector not a point. Well done barometers!

Sometimes I wish I had a weather app that did the same. Open it, and the screen would just say “wear the same as you wore yesterday,” or maybe also wetter/drier/windier.

The app Dark Sky comes closest to that magical feeling, although in a different way: its hyperlocal, to-the-minute forecasts aren’t always accurate, but when it says “rain stopping in 12 minutes” and then, in 12 minutes, the SKIES CLEAR and you can go outside without bothering to take a hat… it’s a superpower.

I read recently that weather forecasts are suffering because flights have been cancelled, and aircraft are responsible for a large amount of the data that goes to feed the simulations:

The European Centre for Medium-Range Weather Forecasts reports a 80% drop in meteorological readings due to cancellations of commercial flights. According to their study, removing all aircraft data from weather models reduces accuracy by 15%.

(You might guess that I took the atmospherics module at uni, and meteorology was my favourite part of geography at school.)


How about hyperlocal, to-the-minute pandemic forecasts?

The UK govt announced its COVID Alert Levels which run from 1-5. Here they are [pdf]. e.g.:

  • Level 2: COVID-19 is present in the UK, but the number of cases and transmission is low
  • Level 4: A COVID-19 epidemic is in general circulation; transmission is high or rising exponentially

Like a barometer, the level takes into account situation and direction of change and rate of change.

The government also released an equation which has been roundly mocked. Here it is:

COVID Alert Level = Rate of infection + Number of infections

Which is… fine? I don’t get the mocking. I mean, it communicates exactly the right information. The alert level goes up if either of the other two numbers go up. What were they supposed to write? It’s impressive wordsmithing to convey that entire mental model so concisely.

But what caught my attention was that if the alert level rises, the lockdown would be once again tightened… and they said this could happen locally.


What would it mean to have a giant pandemic simulation running on those impressive Met Office computers?


Could this pandemic mirror world be used for forecasting?

How many million data points would it need to be fed to forecast

  • how many contagious infections there are in a particular neighbourhood, at a particular time,
  • the effective rate of infection, given crowd levels etc, and whether crowds are likely to form because of the weather and so on,
  • how to compare levels of exposure, just for you, given your route.

What sensors would be required to feed such a simulation?

How fine could the resolution become?

Could some kind of future Dark Sky meets Citymapper meets contact tracing app say things like…

“well you’re near London Bridge and the general number of infections is low, but there’s been an infection wavefront moving up slowly from the south plus, huh, Tuesday morning it usually gets kinda busy, so between 8-10am in that area we’re forecasting a COVID Alert Level of 4.3.

“BUT the surrounding neighbourhood we’re looking at a local alert level of 3.6 and falling,


“if you get off the train one stop early and walk the rest of the way to work, sure you’ll be out in public for 30 minutes longer, but your exposure overall will still be lower, so that’s your recommended route this morning.”


Communicating this might end up looking a bit like rads, the unit of absorbed radiation dose.

Maybe your phone could track your location and give you a live exposure number over the day, like a badge? It’s 2pm and you’re at 40 co-rads today. We recommend you leave before rush hour and take this 20 co-rad route home, also WASH YOUR HANDS.

More posts tagged:
Follow-up posts:

If you enjoyed this post, please consider sharing it by email or on social media. Here’s the link. Thanks, —Matt.