Disclaimer: This post contains some very preliminary thoughts on a topic that I believe would be of interest to some people here. There are probably better expositions on the subject that I haven't been able to find. If you know of such expositions, I'd appreciate being pointed to them.

There are qualitative differences between the types of forecasting that is feasible, or most suitable, for different time horizons. In this post, I discuss some of the possibilities for such time horizons and the forecasts that can be made for those.

The present (today)

Predicting the present doesn't involve prediction so much as it involves measurement. But that doesn't mean it's a slam dunk: one still needs to make a lot of measurements to come up with precise and accurate quantities. One cannot simply count the entire population of a region in one stroke. Doing so requires planning and a detailed infrastructure. And in many cases, it's not possible to measure perfectly, so we measure in part and then use theory (such as sampling theory) to extrapolate from there.

The very near future (tomorrow)

The very near future differs from the present in that it cannot be measured directly, but measuring it is often no more complicated than measuring the present. In a discrete model, it's the next step beyond the present. An example of a tomorrow prediction is: "what restaurants will be open in the city of Chicago tomorrow?" For any restaurant to be open tomorrow, it is most likely either already operating today, or has applied to open tomorrow. In either case, a good stock-taking of the situation today would give a clear idea of what's in store for tomorrow. Another example is when people make projections about employment or GDP based on asking people about their estimated workforce sizes or production levels in the near future.

Predictions about the near future involve a combination of the following:

  • assuming persistence from the present
  • asking people for their intentions and estimates
  • identifying and adjusting for any major sources of difference between today and tomorrow. In the restaurant case, an example of a major source of difference would be if "tomorrow" happened to be a major festival where restaurants customarily closed.

Who forecasts the very near future? As it turns out, a lot of people. I gave examples of economic indicator estimates based on surveys of representative samples of the economy. Also, I believe (I don't have an inside view here) that industry associations and trade journals function this way: they get data from all their members on their production plans, then they pool together the data and publish comprehensive information so that the industry as a whole is well-informed about production plans, and can think a step ahead. (SEMI might be an example).

The near but not very near future, or a few steps down the line

For the future that's a little farther out than tomorrow, simply assuming persistence or asking people isn't good enough. Persistence doesn't work because even though each day is highly correlated to the next, the correlation weakens as we separate the days out more and more. Asking people for their intentions doesn't work because people themselves are reacting to each other. For inanimate systems, different components of the system interact with each other.

This is probably the time horizon where some sort of formal model or computer simulation works best. For instance, weather models for the next 5 days or so perform somewhat better than the fallback options of persistence and climatology, and in the 5-10 day range they perform somewhat but not a lot better than climatology. Beyond 10 days, climatology generally wins.

Similarly, this sort of modeling might work well for estimating GDP changes over two or three quarters, because the model can account for how the changes in one quarter (the very near future) will have ripple effects for another quarter, and then another.

The problem with such models is that they quickly lose coherence. Small variations in initial assumptions, to a level that we cannot hope to measure precisely, start having huge potential ripple effects. Model uncertainty also gets in the way. The range of possibilities is so large that we might as well get to more general long-term models.

What is the value of making such predictions? The case of weather prediction is obvious: predicting extreme weather events saves lives, and even making more mundane predictions can help people plan their outdoor events and travel and can help transportation services better manage their services. Similar predictions in the economic or business realm can also help.

The organizations who specialize in this sort of prediction tend to be the same as the ones predicting the very near future, probably because they have all the data already, and so it's easiest for them to run the relevant models.

The medium-term future

This is the part of the future where general domain-specific phenomena might be useful. In the case of weather, the medium-term future is general climatology: how warm are summers, and how cold are winters? When does a place get rain?

Computer simulations have decohered, and formal models that are sufficiently realistic in the short term get too complicated. So what we do use? General domain-specific phenomena, including information about equilibrating and balancing influences and positive and negative feedback mechanisms. Trend extrapolation, in the (rare?) cases that it's justified. Reality checks based on considerations of the sizes and growth potentials of different industries and markets.

The medium-term future is the time horizon where:

  • New companies can be started
  • City-level transportation systems can be built
  • Companies can make large-scale capital investments in new product lines and begin reaping the profits from them
  • Government policies, such as overhauls to health care legislation or migration policy, can be implemented and their initial effects be seen

My very crude sense is that this is the highest-leverage area for improvements in forecasting capabilities at the current margin. It's far out enough that major preparatory, preventative, and corrective steps can be taken. It's near enough that the results can actually be seen and can be used to incentivize current decision makers. It's far enough that direct simulation or intricate models don't stay coherent, but it's far enough that intuitions derived from present conditions, combined with general domain-specific knowledge, continue to be broadly valid.

The long-term future

The dividing line between the medium-term and long-term future is unclear. One possible way of distinguish between the two is that the medium-term future is heavily grounded in timelines. It's specifically interested in asking what will happen in a particular interval of time, or in when a particular milestone will be achieved. With the long-term future, on the other hand, timelines are too fuzzy to even be useful. Rather, we are interested simply in filling in the details of what it might look like. A discussion of how a world that's 3 degrees celsius warmer, or of space travel, or of a post-singularity world, or of a world that is solar-powered, might fit this "long-term" moniker. Robin Hanson's discussion of long-term growth and the multiple modes of such growth also fits this "long-term" category.

With the long-term future, simply painting futuristic visions, informed by a broad understanding of theory to separate the plausible from the implausible, might be a better bet than reasoning outward from the present moment in time or from the "climatology" of the world today. Indeed, as I noted in my discussion of Megamistakes, there may well be a negative correlation between having a clear vision of the future in that sense and being able to make good timed predictions for the medium term.

With the long term future, are there, or should there be, incentives to be accurate? No. Rather, the incentives may be in the direction of painting plausible (even if improbable) future scenarios with the dual goal of preparing for them or influencing the probability of achieving them. This means dampening the probability of the catastrophic scenarios (even if they're low-probability to begin with) and increasing the probability of, perhaps even directly working towards, the good scenarios. On the good scenario side, a futurist with a rosy vision of the future might write a science fiction or speculative science book that, a generation or two later, inspires an entrepreneur, scientist, or engineer to go build one of those highly futuristic items.

Nick Beckstead's research on the overwheming importance of shaping the far future makes the relevant philosophical arguments.

I could probably split up the long term further. I'm not sure what some natural ways of performing such a split might be, and I also don't think it's relevant for my purposes, because most long-term forecasts are hard to evaluate anyway.

PS: My post on the logarithmic timeline was a result of similar thinking, but they ended up being on different topics. This post is about the qualitative differences between time horizons, that post is about having a standard to compare forecasts for different time intervals in the future.

New Comment