Hacker News Reader

Comment by econperplexed | Hacker News Reader

OT: I see economists of various ideological tilts constantly arguing about how macroeconomic variables will respond to specific government policies. You know: "this new tax will increase inflation; no it will not, it will increase unemployment".

What I don't get is why there are no readily available online macroeconomic simulators with real world data that make these sort of predictions. For a lot of countries, up-to-date macroeconomic and demographic is readily available.

Surely, economists of every ideological tilt have their own standard model of macroeconomics [1] which they use to make rough predictions? If not, how complicated is it really?

[1] Much like the standard model of particle physics or of cosmology. Fairly dirty and complicated models with lots of tunable parameters.

View on HN

>[1] Much like the standard model of particle physics or of cosmology. Fairly dirty and complicated models with lots of tunable parameters.

Anyone capable of making models that can predict the future with a reasonable degree of accuracy is probably in finance, making 5-10x what they could in economics. The economy is an incredibly complex system, subject to emergent behaviour and chaotic effects, much more so than in physics. You'll see papers from the large hadron collider where they prove stuff to within p=0.000001; it's impossible to prove anything to that degree of confidence in economics. Especially because it's not possible to create true controlled experiments: you can't have two otherwise identical societies that differ by just one factor, rather there will always be other ways in which they differ to, and how these are accounted for can have a big effect on the output of models.

Any model has some precision associated with it. I am not asking for 5 sigma level of precision. But if the economists are fighting about what will happen to a particular macroeconomic variable, they are making a prediction even if it is a log_2(3) bit prediction [1]. There must be some math that is backing that. I want to see it explicitly put into a computational model that runs on live data. Otherwise is there any real content to the arguments that economists are having?

> Especially because it's not possible to create true controlled experiments

Cosmology has zero-experiments. It is a completely observational science [2]. And in fact, has very bad data. Basically a time-frozen snapshot of the universe from a particular point in space. They can't even make predictions about the future, only about what some new dataset of that same time-frozen universe will say. Economics on the other hand has the benefit of lots data about interventions and their consequences. There is so much opportunity to develop models by making predictions and checking what happens.

[1] variable goes up or down or stays the same.

[2] The fact that the roots of modern science lie in a purely observational no-experiment discipline of astronomy is lost on many.

> But if the economists are fighting about what will happen to a particular macroeconomic variable, they are making a prediction even if it is a log_2(3) bit prediction [1]. There must be some math that is backing that. I want to see it explicitly put into a computational model that runs on live data

This is what (some) people in finance do: make models to predict things, because if something about the future can be predicted to a sufficient degree of accuracy, it's generally possible to make money from it. In economics, the incentives are slightly different; in academia, the incentives are to publish interesting/novel/topical papers, like with other social sciences, not necessarily to make repeatable predictions. In social science nobody gets punished for making an interesting model that hasn't been rigorously proven to make repeatable predictions, while in finance on average better models make more money and get rewarded more. But sharing an effective model means other people can use the predictions too, meaning you capture less value from the predictions yourself, so people with an effective model have an incentive not to share it

Few scientific hypothesis have been tested by so much money as that there is no alpha in the market. I mean, the amount of money that goes into proving that wrong is perhaps larger than what went into finding the Higgs boson.

Best you can get is some bounds for whatever effect you want to predict and those bounds will open wide very fast.

As for lots of data, doing proper econometric models requires really well gathered data. Data is published quarterly or yearly, so you do have comparable series for international country level macroeconomics, but those are actually short considering how many variables must interact.

> I want to see it explicitly put into a computational model that runs on live data.

I _also_ want to see it. But I do not want to build it, since it seems incredibly boring.

> What I don't get is why there are no readily available online macroeconomic simulators with real world data that make these sort of predictions.

Are you high or what?

The problem is so complex that people don't even agree how to take a measurement, let alone how to simulate the system. If nowadays people disagree when looking at the same numbers, imagine if each one could point to the numbers in his own simulator.

Lets take a very particular variable. Every month or so, central/state banks in every country have a meeting to decide the policy rate. They look at some data and the output of some models, then wave some political magic wand over it and make a decision about which direction to move the rate. Then they make a statement like, "our decision will decrease inflation over the next three quarters by this much."

The data they use is usually publicly available, the models used by these must not be very secret. Is it not possible to hook this all up into a nice package, where I can go in and tune some knobs to my liking (there will always be knobs) to get a prediction on the inflation over the next three quarters.

Then, I can publicly claim, "Given this data, and this model, and my knob settings, here are my prediction." Then some other economist can come and dispute the data, model and the knob settings. They select a different data source, model and their own knobs.

Then three quarters later we check who made the right prediction.

Notice, I am not asking for the correct model of reality. Only a playground where people can argue using concrete data, models and knobs.

The Danish Ministry of Finance has a public model they use to forecast the economic impact of certain policy decisions called MAKRO: https://github.com/DREAM-DK/MAKRO/tree/main

Ok fine. My strong reaction was really due to the use of the word "simulator".

But if you say model - yeah, central banks do that. I'm not really familiar how public their models are. I would bet you that they intentionally don't want people to know "here's the exact model and here's the exact inputs", because they want to keep some discretion for themselves. They want to be able say "I know that this model with these inputs says X, but as a thinking human being I know that the model doesn't account for Y which has been crucial in these past few months, and so instead of X we'll do Z".

What you said makes a lot of sense, but it would require collaboration from the decision makers. There's no incentive, I believe, for someone to do this "open source stye", because the moment that their effort pays off, some hedge fund will take it and use it to trade on it.

The only real rule in economics, other than simple accounting identities, is that people will do what they think is best for themselves. But everyone's utility curve is different, let alone their perception of their utility curve. It's incredibly difficult to rigorously model.

At sufficiently large scale we should be able to model certain personalities though.

Imo a decent economics simulator is indistinguishable from a full physics simulator. There’s no other way to account for the actions of millions of independent actors with unique incentives. Econ models tend to focus on very specific independent variables and hold everything else constant because there are just too many dimensions to consider in a full model

> What I don't get is why there are no readily available online macroeconomic simulators with real world data that make these sort of predictions.

You aren't the first person to have this thought. It just turns out to be incredibly difficult.

Sounds like we need wikiconomy crowd-sourced model building over two open-source platforms:

- mathematical macro-economic simulation run by differential dependencies (like Minsky)

- agent-based Monte-Carlo platform for behavioural experiments to find or validate the macro rules

There would be a core set of models, with parametrized instances running for each country, loosely coupled through trade flows, capital flows, FX markets and migration.

My point is that tons of people have wanted to do this and tried and failed. There is too much to model and existing models don't have predictive power. Basically when calibrating the models there will multiple sets of parameters that fit past data and produce wildly different predictions. If you could devise a general macroeconomic model that made accurate predictions you could become ludicrously wealthy. Even accurately predicting corn prices 5 years out would net you considerable returns.

So many things that effect economic behavior just aren't really quantifiable. Even Keynes who was a proponent of modeling and economic planning acknowledged this[1]. Hayek talked about information being too decentralized and preferences arising organically, which made planning(modeling) imprecise and inefficient if not impossible[2].

[1] https://en.wikipedia.org/wiki/Animal_spirits_(Keynes)

[2] https://www.kysq.org/docs/Hayek_45.pdf

Why do you think economics is as deterministic as particle physics? Why should a simulation be any more informative than an economist's assumptions?

Particle physics isn't that deterministic either. It uses a lot of statistical tricks and hard engineering to reach its results. That's why there are so many physicists working as quants.

Economics is a science and what an economist calls economics is not what the market things an economist knows.

In a science you are interested in testing hypothesis.

Econometric/statistical models will generally serve to test general hypothesis. What happens if we raise interest rates? Will inflation go down? This can be tested in many ways. But it is not a concern to actually predicate the values of macroeconomic variables. It’s really not interesting academically and basically impossible to do right.

This is of course completely out of vogue since ML wiped out the scientific method education of us all.

Because a good model would be like a self-defeating prophecy.

An “economy” is composed of people, so maybe we can do this once we’ve cracked simulating people.

psycohistory!

but but ... you cannot have data get in the way of good theory.

Reply on news.ycombinator.com