economic forecasting

amandstu

I am not sure it is so important (isn't it as good to just "do your best" ?) although avoiding sudden changes is obviously very important.

However my question is as to whether the " weather forecasting model" is used for this purpose?

I mean when you forecast the weather (or the climate) you divide the world into as small as possible regions , take readings from each area and perform milions of parallel calculations to determine a global outcome a short time ahead.This process is repeated into the distance until the outcomes become unreliable (about 5 days I think it is)

Will this work (is it used) with economic (financial) forecasting ?

I suppose you would take an inventory of all the potential spenders (corporate for the most part perhaps but also individuals too )at any one time ,examine their financial options , take a "logical" decision on each one's part and calculate the new "inventory" as a result of the millions of simultaneous actions.

Clearly the process is approximate but would there be any benefit from doing it?Has it been tried? How accurate could it be with enough computing power ?

andrew

The model the ESRI uses to model the Irish economy in the medium term is the HERMES-13 model. You can read about it here, in a paper from July of this year.

amandstu

thanks very much.Be careful what you wish for comes to mind!

I can't promise to go through that in a hurry (I wish I could digest it at the same speed as our intelligence services) -it does look interesting though.

andrew

I don’t know anything about weather modelling, and don’t have any formal background in economic modelling, but I read a bit of the paper.

First, you've got your basic assumptions. The model is built around the basic assumptions that a) firms are attempting to minimise their cost of production or maximise their profits and b) that households are attempting to maximise their utility. I guess these assumptions would be analogous to the basic assumptions which a weather forecasting model makes about molecules/units of the atmosphere. Many other assumptions exist within the model regarding interactions between various elements and the determination of other variables, but these are the most basic ones. Interestingly, the paper notes that “goodness of fit to historical data is sacrificed to ensure that the model’s specification conforms to economic theory on the behaviour of firms and consumers.” I guess the logic is that consistency is more important than trying to estimate parameters which aren't fully understood and can’t (yet) be modelled.

Beyond a reliance on basic assumptions, however, the analogy between Economic and weather modelling starts to break down. The time horizon of the HERMES-13 model is 5 or more years. This is the fundamental difference between Economic models and meteorology models. Economic models, even ‘short run’ ones, model ‘climate.’ Meteorology, models can simulate ‘climate’ (long term) and ‘weather’ (very short term).

There are two fundamental issues here. One relating to data, and one relating to the difficulties of economic modelling itself. There are two problems regarding data. First there is a big lag between when economic activity occurs and when it’s measured. Second, not all economic data is readily measurable. The time lag is at a minimum about 3 months for GDP, and much longer for all other kinds of data. In contrast, it’s possible to get an almost real time picture of the state of the atmosphere via weather stations etc. In terms of measurability, short of installing cameras in everyone’s homes and following them around every day and reading their minds, we can never know all that we’d need to know to create a perfectly accurate model. Some of the data isn't even quantifiable. In contrast, every short run weather variable seems to me to be pretty readily quantifiable and measurable. As such, even if aliens created a perfect model of human behaviour in the short run, we wouldn't have enough data to run the model. So, we’re limited by data to long run models. But even if we weren't limited by data, there would still be problems.
In the short run, a model can ignore lots of complicating factors. In a weather prediction model, you don’t have to take into account the fact that increased carbon dioxide leads to a positive feedback loop which releases more carbon dioxide from formerly frozen Arctic Tundra. You don’t have to focus on long term trends in the heat capacity of ocean water and ocean salinity. If you want to predict next week’s weather, you can focus solely on the behaviour of air and how it relates to other short term inputs, without consideration for longer term trends. This is a huge simplification. Since air is readily modelled (we know exactly how it behaves for a given input), and because we’re only modelling air, and because we have mountains of data regarding the state of the atmosphere at a given time, we can create the short term models you mention. Obviously I’m making some general simplifications here since I know nothing about meteorology. But I think it’s safe to say we've lots of knowledge about how air behaves and we can model it well.

In contrast, if you set about trying to model the day to day behaviour of individual people and firms (by modelling them atomistically as you describe) you’d have a much tougher task. Not only would you not have the required data, but arguably human interactions only can only be modelled accurately in the (relatively) long term anyway. Humans aren't like air particles which react in a very specific and constant way to a given input. In the short term, individuals and firms might make decisions which are very, very difficult to model. It might be that their decisions on a given day have no basis in any kind of thought processes which can be modelled or generalised, and are only ‘economic’ in the long term.

To be sure, attempts have been made to model individuals more accurately. Game theory tries to model individual decision making in the short term, and it’s a pretty good tool for describing how people act when faced with a given decision. However, it doesn't generalise into an entire economy. In addition it’s also unclear as to how applicable Game Theory is to individuals in the real world (this is the measurement problem again). And while Macro models do have ‘micro-foundations’ these are micro insofar as they relate to individuals, not in terms of the time period to which they relate. We can’t explain individual decision making very well at all. And so it might be that we can only ever model the economy in a long-term, economy wide sense, because economic agents are only model-able in the long term.

If this is the case, then it precludes the use of a model such as the one you describe. To the extent that you’d use such a model to predict behaviour into next week etc. by modelling individuals, it couldn't because we have neither the models nor the data. To the extent that you’d use some longer run version of them model, you’d necessarily have to include a lot more than just individuals and firms. You’d have to include structural features of the economy, such as the labour market, as in HERMES-13.

Specifically, HERMES-13 takes a set of inputs (exogenous variables), such as the unemployment rate, the interest rate etc. and then uses several hundred equations in order to model how these input variables affect the dependant (output) variables. These equations are grouped into 26 ‘blocks,’ totalling 804 equations in all. Each block models a specific sub-set of the economy; these blocks I are the ‘structural’ aspects of the model which affect the economy in the longer run. You've got a block for agriculture, for the housing market, for labour supply, for demographics, for energy etc. Maybe these blocks are analogous to a climate model which takes into account things like global warming, and the exact long term interaction between oceans and atmospheric conditions, and long term cycles like El nino etc. Each block transforms the input variables in a certain way, and there are feedbacks between various blocks too. For example, interest rates generate many different kinds of feedbacks. The resulting output variables provide a general picture of how the economy might do in the coming five years, in the same way that climate models predict average temperatures, ±a tenth of a degree or so.

So until economists have the ability to a) model human and firm behaviour accurately on a small scale and short time frame and b) have granular enough data to input into these models, it’ll be difficult to model the economy in the way you describe. We’re stuck with much more general models for now, which appear as though they preclude the approach you describe.

To reiterate, I don’t have any formal background in economic modelling, so please take the above with a pinch of salt.

Scofflaw

andrew wrote:

First, you've got your basic assumptions. The model is built around the basic assumptions that a) firms are attempting to minimise their cost of production or maximise their profits and b) that households are attempting to maximise their utility. I guess these assumptions would be analogous to the basic assumptions which a weather forecasting model makes about molecules/units of the atmosphere. Many other assumptions exist within the model regarding interactions between various elements and the determination of other variables, but these are the most basic ones. Interestingly, the paper notes that “goodness of fit to historical data is sacrificed to ensure that the model’s specification conforms to economic theory on the behaviour of firms and consumers.” I guess the logic is that consistency is more important than trying to estimate parameters which aren't fully understood and can’t (yet) be modelled.

Heh. When you think about it, that's an absolutely extraordinary statement. Climate, and indeed other complex models, are tested by 'back-casting' against historical data. If your climate model can't model historical data, then it's unable to reproduce observed output from known inputs, and its value in forecasting is obviously not great.

What the statement “goodness of fit to historical data is sacrificed to ensure that the model’s specification conforms to economic theory on the behaviour of firms and consumers.” means is that the economic models, rather than being judged on their ability to reproduce known outputs from known inputs, are judged instead on whether they conform to accepted economic theory.

The more I think about it, the more bizarre that gets. If the model has good conformity with economic theory, and in order to achieve that conformity, the model's ability to reproduce reality was sacrificed, then...well, the model is useless as a guide to reality, and tells you that so is economic theory.

Surely some mistake, as PI used to say?

cordially,
Scofflaw

andrew

I know back casting is used for a lot of economic models, so this characteristic of HERMES definitely isn't one that's true of all economic models.

Without knowing exactly which historical trends this model doesn't fit to, it's hard to know exactly what they mean when they say it doesn't fit. Maybe it's the case that there was a structural 'break' in the Irish economy sometime around the beginning of the Celtic tiger, which renders subsequent models (like HERMES-13) incapable of fully describing the Irish economy before this time. Sure they could 'tune' the model to fit the data, but in doing so you don't make the model any better at predicting the future.

Scofflaw

andrew wrote: »

I know back casting is used for a lot of economic models, so this characteristic of HERMES definitely isn't one that's true of all economic models.

Sure - that's what surprised me with this one.

andrew wrote: »

Without knowing exactly which historical trends this model doesn't fit to, it's hard to know exactly what they mean when they say it doesn't fit. Maybe it's the case that there was a structural 'break' in the Irish economy sometime around the beginning of the Celtic tiger, which renders subsequent models (like HERMES-13) incapable of fully describing the Irish economy before this time. Sure they could 'tune' the model to fit the data, but in doing so you don't make the model any better at predicting the future.

Their statement is pretty blunt, and fairly specific - "goodness of fit to historical data is sacrificed to ensure that the model’s specification conforms to economic theory on the behaviour of firms and consumers". If there were a structural break in the Irish economy that the model wasn't able to cope with, that would usually be stated explicitly (as it is in some IMF modelling papers, for example).

cordially,
Scofflaw

andrew

Scofflaw wrote: »

Sure - that's what surprised me with this one.



Their statement is pretty blunt, and fairly specific - "goodness of fit to historical data is sacrificed to ensure that the model’s specification conforms to economic theory on the behaviour of firms and consumers". If there were a structural break in the Irish economy that the model wasn't able to cope with, that would usually be stated explicitly (as it is in some IMF modelling papers, for example).

cordially,
Scofflaw

Looking back at the paper, the full sentence is actually "If necessary, goodness of fit to historical data is sacrificed to ensure that the model’s specification conforms to economic theory on the behaviour of firms and consumers." So it's not entirely clear whether they've actually had to do so.

Either way, I can see their logic. If you have a model which doesn't describe some aspect of the economy, then either a) your assumptions are wrong or b) there's some factor you're not considering. If there's a decent amount of other evidence that the assumptions you make aren't wrong (which there is), then it doesn't make much sense to change them. It probably makes a bit more sense to see whether you're not considering some factors, first.

I often wondered about this... economic forecasting consists of looking at various bits of data and analysing them and coming up with an answer/forecast that is plausible.

A certain amount of it is going to be personal opinion so for example I am an optimist and would always interpret information in an optimistic manner because of my personality, however if you were an economist with a pessimistic outlook on life, then you are going to intrepid the information in a pessimistic manner therefore two economist could look at the same data and come up with different interpretations because of their personalities.