[ExI] Serious topic

[Anders Sandberg]

Keith Henson wrote:
> http://www.ultimax.com/whitepapers/ETP1_ThreeSigns.pdf
>
>   

I'm having serious problems with Hubbert curves, for the same reason I 
have become less happy with using Moore's law for long range prediction. 
Just like Bass technology diffusion curves they fit data very well in 
retrospect. But when used to extrapolate the future they seem too 
unstable: the predicted peaks or sigmoids jump all over the place due to 
noise in the data.

When you have a fairly well developed curve (i.e. you are far beyond the 
hump of a peak or the inflexion point of a sigmoid) extrapolation is 
robust. But before that - at the hump or inflexion point, or even a bit 
after - the extrapolation is almost useless. It is worth testing it 
yourself with a toy model with noisy data, the effect is a bit 
surprising. Those confidence intervals get very wide.

It is better to try to shore things up using other kinds of data or 
modeling. UK oil production is clearly moribound, but it is also a 
particular rather small pocket. Oil reserve data is famously uncertain 
and biased by various interests. It might be more revealing to look into 
how the markets are investing long-term.

-- 
Anders Sandberg,
Future of Humanity Institute 
James Martin 21st Century School 
Philosophy Faculty 
Oxford University