Take Home Test

I’ve always been a big fan of FT/Alphaville — as evidenced by once inhabiting a telepresence bot at their summer event a decade back and heckling a transhumanist speaker from within my temporary technological body, all the way from Manchester, after which I was happy to wheel up to the nearby coffee bar and chat with patrons. A remotely fun time was had.

The FT gang ran a piece today that ties to the most recent post here, The Single Arc in ways that are worth sketching out.

Tl;dr FT/Alphaville: back in January, the European Central Bank (ECB) announced it would ask 110 euro-area banks to write their own doomsday rather than react to one handed to them — a "reverse stress test" where each bank invented whatever geopolitical scenario got it to a 300bps capital hit, then showed the data on how well it fared. Reporter Daniel Davies called it a dystopian collaborative fiction-writing exercise for supervisors, and bet the results would mostly be prosaic rather than wild.

The results are out now. Shock: mostly prosaic, not too dystopian — no asteroids or rogue AI, just recessions and unemployment near historical worst cases (meaning we've been there before). But Davies caught the fun bit: banks routinely wrote themselves a hero's rescue into the scenario, with "mitigating actions" that on average halved the damage.

His diagnosis: the test isn't rigged so much as self-graded, writing "when you're writing your response to a stress test prompt, your main focus is on delivering an answer that you think the regulator will like," which is the opposite of how anyone actually behaves in a real crisis. The banks kept the test within a scenario they could manage, and — surprise— they managed.

This brings to mind critiques of wargaming, where the threat can be as much internal organizational pressure to look good as it is to face down well-armed, clever opponents. There’s the classic (and very timely) story of the Millennium Challenge 2002, a $250 million Pentagon exercise. Lt. Gen. Paul Van Riper commanded the Red Force and sank much of the Blue fleet in the opening hours using (now familiar) asymmetric tactics: small boats, commercial-ship-launched attacks, tactics Blue's sensors weren't built to catch. Rather than let that result stand, the exercise was stopped, the sunk ships "refloated," and Red was ordered not to shoot down incoming aircraft, and to turn on radar specifically so it could be destroyed. That’s an expensive away day.

Another fly in the ECB ointment concerns the fact that the banks in the test were only asked to play out stress tests in isolation. There was reportedly no consideration about the impact of other institutions’ responses to the same stresses in the same market. The supervisors noticed. "Seen in isolation, banks' mitigating actions appear reasonable," the report admits, "but they also need to be viewed through the lens of a potential geopolitical crisis" — because, as it goes on to note, "several banks may attempt to implement specific management actions at the same time, which could impair the effectiveness of those actions."

In other words, no feedback loops, no collective effects on markets, etc. Just Bank X, making good decisions that worked out. Champs. (This latter bit is something we built Foom to explore relevant to decision making around AI — how other big stakeholders’ decisions might shape AI progress, not just those of one company or developer).

I post this because it ties to the issue of treating scenario planning as a “choose your own adventure” exercise in which the scenario most favorable to you is the one you embrace. It doesn’t work like that (or shouldn’t). Stress-tests, simulations and scenario exercises that don’t actually test decision-making, durability or resilience—or at least generate strategic conversations of value—are essentially an expensive storytime.