Climate change flawed statistics

[Cardboard]

https://www.americanthinker.com/blog/2019/10/gift_to_climate_alarmists_exposes_global_warming_junk_science_in_12_minutes.html

 

Whatever you believe in, this author certainly makes solid points on statistics.

[LC]

So first, the actual video is here:

 

 

There is definitely a good point being made but I think the video's author jumps to the 'deception' conclusion too quickly.

 

The point he makes is that by choosing the starting point of a certain time series you can dictate the observable trend.

 

For example if our starting point of a temperature series is when a meteor hit the earth and spiked global temperatures, anything forward looking trend will illustrate "global cooling"

 

Conversely if we elect to use the depths of the ice age as our starting point, any forward looking trend will illustrate "global warming"

 

The question is what is an appropriate point to begin building a model. Some may say, "well lets just use the full data series that we have". This has flaws too.

 

I'll use another example to illustrate this: when building financial models, do we use the full history going back to the 1800s? No we do not. One reason being is that financial crises like 2008 will barely show up. However we need to emphasize that time period because it provides valuable information regarding stressed behavior.

 

Your dataset choice is dictated by your objective. If you are trying to build a generalized model, you want to include time series which exhibit a range of behavior. If you are building a stressed model, you want to include series with stressed behavior, similarly for base-case models you would include benign time series, etc. etc.

 

I do not know why the authors of the national climate assessment report chose the time series they did, but to assume it was for deceptive purposes is an unsubstantiated claim.

[Cardboard]

What is interesting is that every graph presented by so called scientists are ALL using different starting dates and ALL using data points that fit best their thesis. This is not proper use of statistics.

 

If you have ever done any presentation to your boss on a project or financial results you would be hammered in seconds doing such.

 

Same thing when you look at financial results of a company where sales would be compared to profits and to balance sheets but, all on different time frames. You would consider it completely inconclusive.

 

I don't know why the world has to resort to such gimmicks and controversy like this when all they have to say is that we need to lower polution and that oil is a finite resource that we are running out of at reasonable price so let's start to use something else?

[rukawa]

What is interesting is that every graph presented by so called scientists are ALL using different starting dates and ALL using data points that fit best their thesis. This is not proper use of statistics.

 

This assumes there is a proper use of statistics...is there? Statistics used "properly" has produced nothing of lasting value EVER. Its shocking how bad it is. Where statistical models are used in a valuable way e.g. physics, engineering...they don't look anything like the statistics taught by statisticians. For instance, try to find a theory in physics that makes use of regressions equations or factor models.

 

The whole assumption here is that there is some proper way to fit a time series and if we only used proper time series methods as described by the holy brotherhood of statisticians we would be ok. Is there a single example of this ever having been done where it produced something of lasting scientific value?

[Cigarbutt]

^There are few certainties in science (none, some would suggest) and the challenge is how to deal with 'softer' inputs, especially if the inputs are multiple. Statistics is not proof in itself, it's just a tool to put a price tag on uncertainty. The problem is when one tries to move from correlation to causation or from explanation to prediction because one does not automatically means the other.

 

Here's a relevant example where statistical models (including time series) can be used (insurance and reinsurance underwriting and claims management):

https://partnerre.com/opinions_research/rethinking-california-wildfire-risk/

Note: the estimated return period terminology means the time period during which a defined event (and its significance in $) will happen in the future.

 

But your job as the CEO or the active investor is to come up with your own value of risk by using the conclusion derived from the statistics (once the models are evaluated for soundness: right data included, relevant data omitted etc) and combine it with common sense and rational analysis. {For the California wildfires, in addition to the 'climate' variables, a critical variable that was missing from many models was the growing value at risk being built at the wildland interface}. And then come up with the reward, if right; and the consequence if wrong.

Emotions and biases not required. And it's work in progress.

[rukawa]

There are few certainties in science (none, some would suggest) and the challenge is how to deal with 'softer' inputs, especially if the inputs are multiple. Statistics is not proof in itself, it's just a tool to put a price tag on uncertainty. The problem is when one tries to move from correlation to causation or from explanation to prediction because one does not automatically means the other.

 

Probability models can be used to put a price tag on uncertainty...you don't need statistics. Statistical models are generated from data. Probability models can be suggested by data but they often aren't generated by it. There isn't a single regression equation in quantum mechanics. Physicists most definitely use probability models but the probability models emerged from conceptual and philosophical reasoning.

 

Statistical modelling is really about using data to create models. In real science the opposite approach occurs...you make simple assumptions and deduce consequences. Data is used to check the validity of your original assumptions by checking whether the deduced consequences match the data. Data is not used to generate models. Its a very different method and its far more successful and valuable in most cases than the inductive approach statisticians use.

 

 

 

 

[SharperDingaan]

Probability models work, because the results are repeatable. Repeat the test anywhere, and you should get the same result.

The problem is that it is only true in very limited conditions; gravity gets progressively smaller the higher you go; it is only the 'constant' that most people think it is - at 'sea level' - which constantly changes (tides, temp, etc.)

 

Statistical models are limited to the data range. Predictability depends upon how robust the model is, in different data sets of different characteristics. The 'best' one can expect is a useable heuristic. But in today's age of routine distortion, the greater value is to trade against the models prediction.

 

No big deal, as long as NOT heeding the prediction is not catastrophic.

But if the model predicts death, & is only right some of the time; it's best not to engage - as we would otherwise be essentially playing a version of Russian Roulette. If the model were predicting rising sea-level; we would sell the house on the coast, buy the house on the mountain, and just rent on the coast as/when we need it. We would incur an 'insurance' cost, but still be living should the worst happen.

 

... However, we would not only still be living.

A slightly smaller (hopefully) number of people would be crowded (more demand) into a smaller space (less supply), raising the value of all remaining homes. And if our mountain home was now part of the 'new' coast, the value of our home would rise even further. We would be 'anti-fragile', and have benefitted from a bar-bell strategy (Taleb).

 

Point is - the value is in typically doing the opposite of what the 'expert' statistical model says  :)

But if the model predicts death, 'hedge' your position. 

 

SD