The Flaws of Spreadsheet Driven Reasoning
- 3 minutes read - 433 wordsSomewhere around 2019 (!) I read this article: “How the Invention of Spreadsheet Software Unleashed Wall Street on the World” by Janet Burns. Judging by the date I recorded this article in my wiki, I think I must have encountered this during my funemployment after being laid off from WeWork. I had accepted my garden leave from them, had already accepted a job at Bloomberg and was enjoying some Neumann-financed time off.
Regardless, I’d never forgotten some brilliant lines and quotes from it and I re-read the article today.
I was staggered by how well it all worked if you swapped “spreadsheet” for “AI.”
In 2010, a pair of researchers published a controversial economics paper. It was cited by UK politicians to justify austerity measures…[it left] a few rows left out of an equation to average the values in a column
As the piece makes clear, there’s no way to build an intuition for the correctness of the data. There are multiple stories of skeptical managers demanding HP-12C validation of outputs before they’re handed up and these soul-crushing, takeout-and-late-nights-on-Broad-Street, are entirely correct.
“There’s a real problem with spreadsheets…: they have very poor error-checking capabilities…it’s actually very hard to detect—not like with a computer program, where there are tools for checking the software in general.”
The computer scientist Simon Peyton-Jones noted that spreadsheets are the world’s most popular immutable, functional programming languages. Spot on. But a programming language will let you exert more points of test and engineered disaster (“assertions”). A spreadsheet can have error checking, but the richness is necessarily limited. And early spreadsheet jocks had the familiarity and skepticism to never lose sight of this. But as the totemic power of the spreadsheet grew (“stacks of paper would often go unread, but they demonstrated that high-tech calculations had been done”), that skepticism moved from “Garbage In/Garbage Out” to “Garbage In/Gospel Out.”
And the totem becomes self-justifying:
Spreadsheets give this impression of objectivity, a certain accuracy, because they look complicated: there are lots of digits…But they are actually quite vulnerable to small errors, and very brittle, with no mechanism for telling you if they’ve gone off the rails.
Seriously, swap “AI” for “Spreadsheet”.
And finally, management derangement and distance:
Regarding his colleagues who worked on The Big Short, Wolfe commented, “They really make clear that people at the top of financial institutions did not fully understand how instruments designed and used by people under them worked. They made billion-dollar bets without understanding what they were doing. They knew the upsides, but not the downsides.”
We’ll be learning the same lesson again, it seems.