I spent a lot of time doing quant models for financial products. Two lessons:
- It's easy to get wrong. If you see a lot of opportunity all the time, it's your model that's wrong, not the market. Very trivial things can mess up your model, like not understanding what the input data means.
- Good trades are hard to get, common rule of thumb that your boss will tell you. Someone selling a house too cheaply? Either there's lots of other buyers or there's something about the house that you didn't think of.
I remember seeing an FX volatility model that almost never traded. It always said "the market is right, plus minus costs". Now and again it would ding and we'd carefully try to get it done in the market, you wouldn't do like Zillow and surprise all the sellers with a massive payday.
I also wrote an arbitrage bot from Bitstamp to MtGox once. I looked at it, didn't turn it on. Percentage wise it was a massive arb, but you couldn't see this in the raw numbers: credit risk on one of the legs. It just shows you that you still need to understand how things actually work. The model is only a calculator for quantifying opportunities that you understand.
This is perhaps the oddest thing about this story. Zillow must have run into several situations where they were paying more than what's sensible, and their staff must have reported this? Surely at the start of an algo buying program, you are vigilant to evidence that the program is wrong?
> Percentage wise it was a massive arb, but you couldn't see this in the raw numbers...
Counterparty risk, that is where the profit came from. I did the same thing, but went into it knowing what the exposure was and how to mitigate it: don't leave crypto in an exchange's hot wallet any longer than it takes to execute a trade and take profit only on exchanges that you could legally pursue in the event that they fail to execute a USD transfer. I also anticipated the debanking that followed... as far as I know I'm still blacklisted by one bank and two money transfer services. What I didn't anticipate was how many exchanges would get hacked and what that would look like to anybody who aggregated the transactions: I get cold called by actual financial institutions a few times a year, always looking to bulk up their dark pools - they somehow have it in their heads that I'm sitting on billions of dollars in BTC. They likely don't have enough of the puzzle to put together the fact that arbitrage doesn't take much when you only need three confirmation blocks - USD wire transfers were the bottleneck.
The issue with that arbitrage is that it was always one-sided. If you're arbing 2 equivalent markets (e.g. name trading on two different US exchanges), you'd expect to buy and sell on both markets in equal amounts, on average.
Bitstamp vs Mt Gox arb required traders to always buy BTC on Mt Gox and sell on Bitstamp (which is a strong indicator of cp risk). This makes inventory management very difficult - especially towards the end of Mt Gox where fund withdrawal times were of the order of months.
- It's easy to get wrong. If you see a lot of opportunity all the time, it's your model that's wrong, not the market. Very trivial things can mess up your model, like not understanding what the input data means.
- Good trades are hard to get, common rule of thumb that your boss will tell you. Someone selling a house too cheaply? Either there's lots of other buyers or there's something about the house that you didn't think of.
I remember seeing an FX volatility model that almost never traded. It always said "the market is right, plus minus costs". Now and again it would ding and we'd carefully try to get it done in the market, you wouldn't do like Zillow and surprise all the sellers with a massive payday.
I also wrote an arbitrage bot from Bitstamp to MtGox once. I looked at it, didn't turn it on. Percentage wise it was a massive arb, but you couldn't see this in the raw numbers: credit risk on one of the legs. It just shows you that you still need to understand how things actually work. The model is only a calculator for quantifying opportunities that you understand.
This is perhaps the oddest thing about this story. Zillow must have run into several situations where they were paying more than what's sensible, and their staff must have reported this? Surely at the start of an algo buying program, you are vigilant to evidence that the program is wrong?