Read on Twitter
Learning curves are a popular model of tech progress. Basically they say doubling cumulative production -> consistent price declines. E.g., every doubling of total installed solar capacity leads to a 20% reduction in costs. Do they really work? A thread.
The figure above illustrates a learning curve. The axes are log-scale so a linear slope implies a proportional increase on the x-axis (cumulative output) is always associated with the same proportional decrease on the y-axis (unit costs). You see these for lots of industries.
If learning curves "work" they imply you can lower costs quick by rapidly scaling up production of existing technologies today.
Unfortunately, these figures don't really prove anything. Suppose that solar panel cost (and price) falls 11.5% every year (nothing to do with experience), and that solar panel demand rises by 2.2% for every 1% decrease in price. That gives you the following figure:
It's the same log-linear slope. So while a learning curve does produce a figure with a linear slope when you plot log costs and log cumulative output, so does just assuming constant progress and a specific demand curve. (and alternative assumptions get you pretty similar results)
(Why might progress just be constant each year? Maybe it's driven by scientific or technical advances in other fields that do not respond to production in your industry, or maybe it's driven by R&D which can only advance so fast)
The trouble is, under the alternative assumption that progress is just constant each year, then there's no way to lower costs by scaling up production fast. You just have to wait.
Fortunately, we do have some other reasons to think learning curves might be a real phenomenon.
Fortunately, we do have some other reasons to think learning curves might be a real phenomenon.
We can make theories of technological change, where they are like an evolutionary process. Under the right sets of assumptions, they can give you learning curves. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1461801
These models think of technology as collections of subprocesses. During production, people tinker with different subprocesses, occasionally stumbling on ones that improve productivity of the overall technology. These get retained, but they get harder to find as production goes on
In a detailed study of one year in an auto manufacturing plant, Levitt, List, and Syverson document these kinds of changes; they also get a learning curve. https://www.jstor.org/stable/10.1086/671137?seq=1#metadata_info_tab_contents
If we really wanted to see if learning curves were real, a nice experiment would be to massively ramp up production and see what happens to cost. We would want to vary how much production changed over time and across these industries.
We kind of did something like this during World War II! A new working paper by @FrancoisLafond4, Greenwald, and Farmer take a look at World War II using lots of original data. https://www.inet.ox.ac.uk/publications/no-2020-02-can-stimulating-demand-drive-costs-down-world-war-ii-as-a-natural-experiment/
They find that cost reductions associated specifically with experience account for 40-67% of the reduction in costs (measured various ways) during WWII. The rest was due to the passage of time (or whatever it is that is proxying for)
So, do learning curves work? I think they have an element of truth in them. Technologies are easier to improve at first, and then get harder and harder to improve (given a fixed platform), at least on average. More production can accelerate the process of finding improvements.
But historical correlations between cost reductions and cumulative output are too confounded by changing demand to be much use for forecasting the likely effect of a policy change.
In one of the few settings where we can estimate the effect of experience with any precision (WWII), historical rates would have led to significant over-estimates.
Thanks for reading! This thread is based on this newsletter, which goes into more detail on all the above: https://mattsclancy.substack.com/p/how-useful-are-learning-curves-really
PS: if you're an academic, do you mind answering this poll, which I want to use in the next newsletter? https://twitter.com/mattsclancy/status/1351560834565218314?s=20
