My work on this is practical mostly but have also written a more formal paper. Simple mean-reversion algos did extremely well in past year. For example my PSI5 strategy that is based on textbook probability theory gained net 15.4% in 2016 with <5% maximum drawdown. This is one of four mean-reversion strategies I use.

These strategies seem to work well because risk is not directly controlled except in the allocation level (no stops.) As soon as stops are introduced, the profitability decreases.

From years of back testing literally thousands of types of non fundamental based strategies using time based, tick based, unique bar chart type based I have found mean reversion may work great for a while but then will have great drawdowns making it basically unprofitable in the long run. My back tests were done using all high volume futures and Forex going back to year 2000 . Some of the ways that you may see better results with mean reversion is using it inside a Bollinger channel as opposed to a Keltner channel . Any sort of custom indicator that has a upper sell zone and a lower buy zone may be used. My own research has shown mean reversion of little value by itself. One of the main problems that kills profitability with mean reversion on lower time frames is commission.

Might be worth reading up on behavioural finance related coverage on mean reversion, I don't have any particular literature as reference but that field of knowledge might provide some perspective.

You might have a look at:

https://www.quantopian.com/posts/comparing-olps-algorithms-olmar-up-et-al-dot-on-etfs

I don't have any real-world experience with it, but my sense from lots of backtesting is that if you are long-only, over the long run, mean reversion works. However, when the market turns down, there is no protection.

I've been working to "hand craft" a strategy that has a short-term directional indicator and long-short mean reversion trading. So far, I'd say that the long-term (10-years or more) Sharpe ratio is limited to 1.0 max, based on backtesting. So, the question is, what would it buy? Could the same performance be achieved by a simpler ETF strategy, passive or active/passive?

I'm not sure what kind of answer your professor is expecting, but you can probably work up an answer on Quantopian by defining a measure of "market stress", and then looking some measures of reversion across a broad universe of stocks (e.g. the Quantopian Q500/Q1500). If you look at pyfolio (a.k.a. tear sheet), you'll see that recent market stress events are defined there (you could ask Quantopian how the picked them).

Of course, your professor will probably ask you to dig deeper, and consider the underlying fundamentals of the response of the market to stress. As suggested by Jon above, you could look into behavioral finance (e.g. https://en.wikipedia.org/wiki/Daniel_Kahneman ). It is pretty intuitive, though, since if you've experienced it, the "irrational exuberance" (e.g. dot-com/telecom and housing bubbles) and eventual "FUD" in the market are tangible.

Any model that you construct will be based on assumptions. In the case of everybody's favorite statistical arbitrage strategy, pairs trading, Antony is right in saying that cointegration is the key concept. We have this assumption that the elements of the pair that we are trading can be combined in some way to make them stationary, and this stationarity is what makes the whole thing go. When a time series (or linear combination of time series) is stationary, this is just about equivalent to saying that all data points from that time series are drawn from the same probability distribution, with the same parameters across time.

In assuming that a pair is cointegrated, we assume that we know the probability distribution and its parameters. Thus, any behavior that is sufficiently deviant (e.g. outside of one standard deviation) can be assumed to be a one-off deal, allowing us to conclude that there will be a reversion effect. In reality, things are a bit trickier since all of our assumptions of cointegration are based on historical data and pairs may go out of cointegration. The statistical theory is sound, however, so it comes down to how good our detection is for when a pair breaks down (maybe incorporating some sort of momentum measure) and how resilient our overall portfolio is (i.e. we want lots of pairs).

If you want to read more about this, check out the Integration, Cointengration, and Stationarity and Introduction to Pairs Trading lectures on the lectures page (quantopian.com/lectures).