How should we talk about nowcasting horizons?

1

This is one for the whole nowcasting community, I’m very interested in your ideas and opinions. :smiley:

The question is as follows: what wording should we use to distinguish nowcasts at different horizons? In forecasting, this is easy:

  • a short-term forecast / forecast with a short horizon is a forecast for only a few days into the future starting from the current “now” (this is the present in real-time forecasting, or more generally the date until data is available to the model)
  • a long-term forecast / forecast with a long horizon is a forecast many days into the future.

Of course, the actual horizons associated with these terms are context-specific and need to be defined for a given application, but everyone gets the idea intuitively.

In nowcasting however, things are more difficult. What we here mean with the “horizon” is actually the delay with which the nowcasts are made. A nowcast with a delay of zero days is a nowcast for the “present”, a nowcast with a delay of 1 days is a nowcast for yesterday, and so on. Hence, as the delay gets larger, we go further into the past.

So, how would you loosely distinguish nowcasts with short delays vs. with long delays?

  • Short-term nowcasts vs. long-term nowcasts? Couldn’t this mislead people into interpreting this analogously to the forecasting terminology and make them believe that short-term nowcasts are “easier” than long-term nowcasts (the opposite is true of course)?
  • The other way round: Long-term nowcasts for short delays vs. short-term nowcasts for long delays? This then has the same direction in time as in forecasting, but it potentially very confusing?
  • Nowcasts for dates close to the present vs. further in the past?
  • Simply short-delay vs long-delay nowcasts? Does this solve the problem of misinterpretation?
  • …

Something to keep in mind here is that we might also combine nowcasting and forecasting in one tool/model (see e.g. Forecasting in epinowcast), so it becomes even more important to have a consistent and maximally inuitive terminology.

What are your intuitions when your hear about these terms, and can you think of any better terms? As far as I am aware there is not yet an established terminology for this in nowcasting, but please correct me should I be wrong!

2

Hi @adrianlison, it’s great to hear you talking about this, I agree it can be really unintuitive.

To add another option: we could use “lag” where we might now use “delay”. It’s simple and I think of “lag” as complementary to using “lead” for forecasting. So nowcasts are a short or long “lag” and forecasts are a short or long “lead”.

This would mean we could combine nowcast/forecasting and still express the result clearly and consistently, e.g. a prediction with a 3-day-lag (nowcast) and a 7-day-lead (forecast). It also avoids using “horizon” which IMO doesn’t work for nowcasting - instead, if needed could replace that with “period” / “interval” / “window”.

Looking forward to seeing everyone’s thoughts and ideas!

3 Likes
3

@kathsherratt Oh, I think I like this a lot. Also helps to avoid confusion between reporting delays and nowcasting lags.

1 Like
4

Nowcasts for dates close to the present vs. further in the past?

This is how I have been thinking about it. I also like this approach or really the horizon approach more generally as it allows a seamless transition between forecasting and nowcasting which seems important.

Something to keep in mind here is that we might also combine nowcasting and forecasting in one tool/model (see e.g. Forecasting in epinowcast), so it becomes even more important to have a consistent and maximally inuitive terminology.

Ha yes exactly.

I think you are right and there isn’t consistent terminology for this. Perhaps we should talk about this at the next meeting (which is in a few weeks?).

5

Circling back to this we could include something like this in Thoughts on a Ten Simple rule paper - #22 by samabbott

1 Like