3 ms·
Real world machine learning at Jane Street
- huac 9y agoThis is a good post! > An example: modeling the difference in price between an ETF and its basket as a Gaussian would be a mistake, since there are well-defined arbitrage bounds on this difference. The infinite tails of the normal distribution always strike where you least expect :)
- totalZero 9y agoEven if a Gaussian isn't the right model for this difference, I don't think that example always holds. Sometimes the ETF trades at a different time than its basket (the same can be true for ADRs), or there is a substantial immediate dislocation in an individual basket component due to a news headline that doesn't get reflected in the ETF price until a few moments later. Also, you may have situations where disadvantageous carry rates cause the basket-vs-ETF carry trade to either trade wide, or to be ambiguous even when an arbitrage is possible (ie, you would make money if everything stays put but you don't actually feel confident about the stability of the carry rates). For options, even an exemplary arbitrage like put-call parity has situations where dividends and/or forward rates justify substantial disparity. I see how it would make sense for an algorithmic trading strategy to make simplifying assumptions, but within every such assumption is a neglected opportunity, however small or rare, to either make or save money.