I don't really understand why you'd say this. They absolutely have predictive power. This is like saying that because there are error margins that we can't take the mean and extract meaningful data from it.
If that were the case, nearly all classic statistics would be powerless. Spoiler alert: it is not.
It's because the model results project more snow and less snow. More rain and less rain. More severe weather and less severe weather. The example cited by the GP illustrates this - 2 years ago during the drought, climate models predicted continuing severe droughts for California. Now, they predict continuing increased precipitation. The results haven't actually changed, just different subsets are being reported.
The global results are better than the regional results, but tend to run hotter than observations, suggesting they don't yet account for the full range of natural variability.
Statistics don't help you predict a coin flip (actually I think tails is ever-so-slightly more likely)
Edit: the point is not that a particular model can't make predictions. Some models will actually be correct. We just don't know which regional models are correct, and because of the range of results, they cannot all be correct. Therefore, as a whole, the regional models have no predictive power.
The stock market is a good example. Lots of people make lots of money betting on their models, but lots of people lose lots of money when their model stops working. There is also a general rising trend underlying the whole system.
The model that predicts your lifespan predicts people with more and less life, with varying probability. It actually does have predictive power for your lifespan, but it does so in a probabilistic way.
The idea that a distribution doesn't have power because it is a distribution is a bizarre opinion to encounter in a community that prides itself on education, computer science, business acumen, machine learning, and informed decision making.
> Statistics don't help you predict a coin flip (actually I think tails is ever-so-slightly more likely)
Uh... yes it does.
We have laws of large number and expected value. You just have to do a binomial distribution and use the data to fit your model.
Statistical learning is a thing. Element of statistical learning is a classic made by several famous statisticians who are responsible for LASSO, RIDGE, Elasticnet regressions.
Bayesian network is also a thing. Likewise with Bayesian Hierarchical Modeling (HM).
A coin flip is actually the easiest thing to predict and for advance statistical learning model such as HM they start out with Binomial Distribution and usually the example is a coin flip.
Are you seriously asserting that climate models are as direct an application of basic statistics as a coin flip?
Statistics work really well when we understand the system, the initial state, inputs, and can accurately measure the results.
None of those apply to climate models, especially regional models.
Statistics make it easy to mislead yourself that the precision of your numbers implies correctness - like the joke, How do you know economists have a sense of humor? They use decimal points.
> Are you seriously asserting that climate models are as direct an application of basic statistics as a coin flip?
No. Never did.
> Statistics work really well when we understand the system, the initial state, inputs, and can accurately measure the results.
... This is actually the opposite of the truth. When we understand the system fully and can fully enumerate it's interal states and every relevant input, that's when there is no point in statistical modeling.
I really think you fundamentally misunderstood your statatistics education.
If you don't know the system and the range of inputs you don't know what to model. If you can't measure the outputs, you can't validate your statistical model.
The predictions for the current solar cycle are a good example - the predictions were drastically off, the actual values outside the predicted error ranges from the statistical models. It was expected to be a more active solar cycle, based on our understanding of the system, its inputs, and the state from the previous cycle.
Because we didn't (don't) understand the system, our models were wrong.
If your model is not right, leaving out inputs, having the wrong range of variables, etc., it doesn't matter how good your math and statistics are, your model works...until it doesn't.
Then you get to do science and generate a new hypothesis, model that, and try again!
What you can't do is say in 2011, "Models show California will be drier because of global warming" then in 2017, "Models show California will be wetter because of global warming" and claim the models (both?!) have predictive power.
"This will not do. You never will be able to make both of them good for any thing. Take your choice, but you must be satisfied with only one. There is but such a quantity of merit between them; just enough to make one good sort of [model]" --Elizabeth Bennett
(And no, there's not nearly enough "drier" area in the new model to be compatible with the older predictions.)
If that were the case, nearly all classic statistics would be powerless. Spoiler alert: it is not.