The training seems to cover some of the basics of time-series econometrics well, but the signals for momentum, mean-reversion, etc, are truly off. The ribbon methods are not on, the physics momentum method is not on. Where did EWMA of returns or differences of EWMAs of prices get left behind. There is an active academic literature on momentum and it is, unfortunately, not being presented.
Some of the methods presented (ribbons? crossing moving averages, etc?) are not really presented well and many trend estimation techniques (a regression in time? really now!) are bound to fail out of sample. Some really just miss the point. Kalman filters are interesting, but why do you include a secular trend? Why not just local levels and local trend models? (i.e., simple MA(1) models are interesting enough and have some good properties).
Financial markets require, more than anything time-series econometrics or signal processing. SDEs don't help much in algo trading, although they produce some good toy models.
Where are the data snooping (p-hacking) tests? Adjusted Sharpe Ratios? Bootstraps....
This is a standard problem of libraries like Zipline and platform like Quantopian that is beautiful from an IT p.o.v., but seem to neglect the underlying statistics. Computer Scientists (including a lot of the ML crowd) and mathematicians (especially those who can do probability and SDEs) don't know that well about much of the modern stats that is needed. It's seriously hard stuff.