The insurance may be the lowest for self-driving cars with a driver ready to take the controls. DLS

Yes, but will there also become mandatory driverless control of vehicles further down the road, especially in cities to 'control' traffic flow?

People should really watch the history of aviation (or more precisely, the accidents of aviation over the last few decades) In the 70's there were approx 2,000 accidents per year and now it's approx 50. At the beginning, there were a lot of mechanical and engineering design problems (much like the "O" Ring for Challenger), and proportionally speaking fewer pilot error. Then somewhere along the line, that proportion started to shift, when they refined all the electronic and mechanical system of the planes, and pilot error became more and more prevalent. (Pilots and ATC trying to override built in systems have led to disasters). So this will likely be the case for self driving cars. I can almost bet that the insurance rate for self driving cars alone will be much much lower than human drivers. As for DLS's comment if a ready to take control driver along with self driving algos would be the lowest, I would think we have to define who is that driver. If it's the average driver, than still the refined algo is still way better. (Again, we see that the top top top pilots did in some cases avert disaster when systems malfunctioned.) But from my daily commute, I can assure you that the average driver is no Lewis Hamilton.

Mark Dow, is that you?

Completely agree, Eduardo.

...or a whisper so quiet?

One point I'd suggest is important to understand is the tradeoff between the parametrization of a model and the volume of data needed to identify "truth." Traditional statistical models (more specifically, parametric models) impose relationships between the variables. Those asserted relationships, derived from theory, allow an analyst to get an answer from only a little bit of data. That's a good outcome if your assertions are correct (in other words, your theory is reliable). It's a bad outcome if not. Machine learning algorithms in which there is "no model" (more specifically, non-parametric models, or no theory) are free from any imposed constraints on the relationships between variables. However, people don't appreciate just how much data is needed to compensate for that freedom. It's vastly more than most people understand. If you pipe too little data into a machine learning model, the software will still spit out some answer. It's just likely to be very wrong. The fields in which machine learning has a promising future are those that generate nearly-unimaginable amounts of data and the existing theory is bunk. In those situations, traditional statistics don't offer much value, so why not trust the black box? However, in fields in which data production is relatively slow (like medicine) or the existing theory is relatively robust, then the value of machine learning is limited. Personally, I'm not informed enough to have an opinion as to whether finance is a good candidate to be meaningfully disrupted by machine learning strategies or not. At first blush it seems obvious that it should be, but I could imagine counter arguments, too.

Often, the difference between the machine learning approach on one hand, and traditional knowledge engingeering, or even statistical methods, boils down to whether you need the system to explain its reasoning. With machine learning, the system learns directly from data and typically cannot explain to a human user why it reached a particular decision. Knowledge based systems use a representation of knowledge, often formal logic or probability theory, that can be used to yield an explanation. So the suitability of machine learning in finance would boil down to the question: would you be happy to let your machine make trading decisions that you could not fathom and would you let it run even if it made trades that make no sense to you?