Hacker News Reader

Comment by ThrustVectoring | Hacker News Reader

The 2008 crisis wasn't caused by subprime lending. It was caused by financial mis-modeling that caused various financiers of sub-prime loans to grossly mis-estimate the risk they were taking, resulting in having an insufficient amount of capital to back the economic risk they were taking.

Specifically, the models used assumed that the economic risk of a mortgage default in Connecticut was uncorrelated with the economic risk of a mortgage default in Texas. When housing prices are rising nationally, this assumption is basically true - the house can be sold by the bank for what's owed on the mortgage, refinancing is available for balloon payments, and any one borrower's cash flow situation is idiosyncratic. But when housing prices started to fall, this lack of correlation broke down, and many financial institutions became unexpectedly responsible for the losses in this asset class.

View on HN

This might be too tangential, but this reminds me a lot of 538's model of the 2016 presidential election. They had a much higher chance of a Trump victory just before the election (30% vs 1%), and the major difference was that they didn't treat each poll's margin of error as an independent event, they modeled it so that any error in the national poll had a large effect on _every_ state poll.

Yeah the cases have some striking similarity. There's generally two sources of error - bias, and noise. When you combine many measurements, only bias accumulates. Noisy errors in one state poll or one mortgage-based cash flow will be offset by noisy errors in another.

If you get surprised by a change from a low-bias regime to a high-bias one, your old assumptions will lead you to be overconfident in your model. For 538, this just meant publishing numbers that were way too confident. But in the financial context, overconfident modeling means you think you can use more leverage than you actually can.

> For 538, this just meant publishing numbers that were way too confident.

This isn't the case for the 2016 elections. 538 ended up having a much more robust forecast because they accounted for systematic polling error (bias).

https://fivethirtyeight.com/features/why-fivethirtyeight-gav...

>For 538, this just meant publishing numbers that were way too confident.

538 was just about the only polling aggregrator that WASN'T guilty of this during 2016 and were relatively bullish on Trump.

Their mistake during 2016 was releasing a series of editorials not really based on data at all that predicted Trump would eventually peak in support and the remaining voters would rally behind a sensible seeming conservative establishment choice like Marco Rubio. Their theory was plausible and completely wrong. Yet their numbers consistently predicted Trump victory in the primsries as he was an eternal frontrunner for virtually the entire primary.

The problem with the risk model was is that all historical data showed nationwide price decline as impossible/never having happened....

Whenever historical growth rates have always been positive, markets react in shock at the "black swan" event when a 30, 50 or 100 year pattern is all of a sudden broken.

Really? House prices never fell before across the whole US? Not even in the 1930s or various recessions in the 60s/70s?

People don't dump houses when the market drops. Few people sell their house when they're underwater. People tend to live in them until prices come back up. Prices don't drop. Demand drops.

Unfortunately, the balloon payments made this impossible for most.

Reply on news.ycombinator.com