> Treat the markets as a complex dynamical system and use the tools from statistical physics such as the Gibbs Ensemble, to derive internal states from input and output.
No. This is what people like LTCM believe. It does not work, the underlying processes driving markets constantly change.
> - Treat the markets as an encryption algorithm and use the tools from cryptanalysis, such as differential cryptanalysis: Even when unable to decipher the full algorithm (total break), one may still derive details and a subset of system functionality.
- They were probably the first to heavily use Hidden Markov Models (see Baum–Welch algorithm and the IBM speech recognition recruitment) and keep on the frontline with new machine learning algorithms (their deep learning revolution would have started 10-15 years before industry).
Yes, and as a fun note, Peter Brown, their current CEO, was Geoff Hinton's grad student.
- They'd have an extremely solid backtesting pipeline, where any new feature can be stress-tested for signal. Features could be very arcane (% of mentions of the currency on neighboring state television) and are constantly (re-)added and removed: concept drift and market competition would gradually weaken signals, but fresh signals are added to keep the performance.
- The extreme returns for 2007/2008 could be due to the increase in volatility of the crisis (you can make more money when there is a lot of action, and competitors suffer from human herd bias / hysteria), but also, in part, due to them being the first to effectively exploit signals in growing social media platforms and search engines. A few years later it was public knowledge that gauging frequency and sentiment on Twitter was once a valuable signal.
- The NSA/CIA type recruits would not work on industrial spying, but on cryptanalysis, (graph) data mining, OSINT, HUMINT, IMINT, and for the security of the firm (which probably runs a tighter security than the intelligence agencies of smaller countries).
All correct.