Comment by narratives | Hacker News Reader

Comment by narratives | Hacker News Reader

Parent Full thread

Curious if anyone here has evaluated any of these models? I’ve done a cursory search and haven’t found a source to play with. I’m curious to know what variables were included, what variables were excluded, how the future is being predicted, and to what degree the creators any of these models have been accurate in predicting the future previously.

Does anyone have a good explanation as to why it’s prudent to attempt to predict anything about anything 80 years in the future?

Also curious if there is evidence of predicting the future in any field. If so, I’m wondering what the accuracy rate is one month out, one year out, five years out, etc.

I’ve done some modeling with sales data, trying to predict demand in a given market based on a series of variables. The issue is how much weight to put on each variable. With lots of market research and historical data, we were able to get decent estimates a couple of months in the future. Beyond that? It was complete speculation. Especially if we missed an important variable, which is common, since its impossible to know what variables will (and won’t) be relevant in the future.

I imagine climate modeling is more complex than my sales models, meaning more variables that were included, more variables that were excluded (especially the ones that are unknown unknowns), and a much larger margin for error (especially as time increases).

I raise this as I haven’t seen any push back from the HN community on articles like this. Not sure if I’m missing something, if it’s political bias, or a fear of ostracized for questioning these reports. Either way, I’d be fascinated to see some deeper analysis on the topic from the minds on this forum.

Have humans ever created a complex model that has enabled us to predict the future with any degree of accuracy? If so, how far into the future and with what degree of accuracy? If not, what is the purpose of a model extrapolated out to 2100?

Trying to imagine when technology will reach X milestone (AGI, for example) reminds me that it’s probably wise to be less sure of our predictions. The last US presidential election reminded me of this too.

The DOE Energy Exascale Earth System Model is on GitHub: https://github.com/E3SM-Project/E3SM

Here is some analysis of how the predictions made by some climate models have done: https://www.carbonbrief.org/analysis-how-well-have-climate-m...

They have done a remarkable job since the 1970s.

Unpopular opinion warning: Not an expert here but I worked on NASA's "Mission to Planet Earth" program for a few years. Cherry-picking of models that produce results that meet the current agenda is not uncommon. Just as researchers have a (deserved IMHO) reputation for publishing data that concurs with their topics...

https://journals.plos.org/plosone/article?id=10.1371/journal... [...]" found 64% of surveyed researchers reported they had at least once failed to report results because they were not statistically significant (cherry picking); 42% had collected more data after inspecting whether results were statistically significant (a form of p hacking) and 51% had reported an unexpected finding as though it had been hypothesised from the start (HARKing)."

https://www.env-econ.net/2017/07/cherry-picking-results.html

https://ori.hhs.gov/videos/case-study-list/3037

https://www.pnas.org/content/pnas/early/2018/03/08/171075511...

Where to start?

Why attempt to predict future conditions decades out? We do it all the time even though we absolutely know our predictions will not be exactly right. A prediction, even if imperfect, is useful if the prediction contains some bounds or probability estimates. People individually and society collectively makes best estimate judgements based on imperfect data every day, all day.

You've done some modeling of sales data, and the difficulty of predicting human whims leads you to suspect that modeling physics is an impossible and useless task? "I imagine climate modeling is more complex than my sales models" is an understatement of supreme folly. Even if you are a genius, the amount of work invested by tens of thousands of researchers spending their lives on this problem is incomparably more robust that some sales data modeling you tossed off.

> Have humans ever created a complex model that has enabled us to predict the future with any degree of accuracy?

Eclipses and solar positions can be computed to extreme accuracy centuries out. There is no closed solution -- it is all modeled and computed numerically.

The IPCC reports do get revised upward just about every release because the estimates are always conservative. Despite what is widely claimed by doubters, the IPCC reports have not been alarmist, and we can go back 30 years now.

Go back 40 years to Exxon's own modeling of climate change and considering the limited amount of data as compared to today, their estimates are not far off.

https://www.theguardian.com/environment/climate-consensus-97...

Outside of astronomy, I don't think there's any history of reliable 80+ year scientific forecasts that predict specific changes to our world.

There’s so much modeling that couldn’t even be done 80 years ago. Electronic computers were in their infancy, climate science was just beginning, weather satellites (hell any artificial satellite) had not yet been launched.

The inputs, the models themselves, and the main way to compute an output had mostly not been invented.

This is a fundamentally unfair comparison. If it couldn’t be computed analytically, it wasn't really possible to compute it 80 years ago. That absolutely doesn’t mean we should just ignore these predictions. Are we supposed to always wait some arbitrary time before listening to long term prediction?

We've had mathematical models of change for over two thousand years. We have reliable, precise ones for astrological changes. As far as I know, we have none related to ecology or other complex systems that have proven capable of making long-term predictions.

I never said we should ignore any prediction. My calculation is purely Bayesian: how much prior trust should you place in a model (or class of models) with no long-term track record of success?

Reply on news.ycombinator.com