Sally,

When you heavily weight portfolio construction by maximizing Sharpe ratios, you short volatility.

Could you expand on that notion? Any maths/papers/graphs would be beneficial to understanding this better.

If you used beta-adjusted returns the Sharpe of the portfolio might be more 'correct'?

I wish I could James, but it is my own observation and not supported by a lot of academic research. You might imagine the survivor-ship bias any study of actual fund performance might have, never mind the difficulty getting data on algorithmic traders who went bankrupt shorting volatility and the sparsity of statistics. Instead, it is what I have observed over my 30 years in algorithmic trading having witnessed the 1987 crash, the 1997 Asian financial crisis, the 1998 Russian stock market collapse, the 2000 .com implosion, the 2001 terrorist attacks, the 2007 and 2008 crashes and the dozens of much smaller, more contained v-shaped events. In 2011, Goldman Sachs unwound a long-short equity strategy where many fundamental quant strategies experienced a 10% move intraday. Afterwards, you could see the firms that "blinked" because they reacted too quickly to an anomaly and shut down their funds, whereas funds that had confidence in their strategies withstood the pressure and recovered and more importantly stayed in business.

Quantopian advertises the black-box nature of its clients algorithms removing any chance of distinguishing irrationality from a broken strategy.

For now it is an empirical question that Quantopian should be able answer. Take the hundreds of algorithms that are in the contest, take their rankings (or just their Sharpe ratios) and produce a scatter plot with one-minute, one-hour, one-day, five-day, etc. returns. Let's see what the result is.

I asked Dan to provide the winners' daily returns, but a better analysis might be to take the daily returns of all of the submitted algorithms, the daily statistics (rank, Sharpe ratio, etc.) and analyze the correlations especially over the past week or so and compared to some volatility measure over the longer term.

Google shows many papers on beta-adjusted Sharpe ratios.

Have to agree with Sally on the over-reliance on a Sharpe ratio, particularly with respect to upside volatility. As far as I can tell, Sharpe ratios do not distinguish between upside and downside volatility (Which may be your point if I'm understanding "shorting volatility" correctly). This hurts trend-following strategies (with respect to mean-reverting strategies) in particular. I suspect plenty of the latter got burned yesterday.

A cautionary quote from Jerry Parker:
"Mean reversion works almost all the time, and then it stops and you are kind of out of business. So the market is always reverting to the mean except when it doesn’t, and then we [trend followers] are on board this big huge trend and these people lose a lot of money. There are two competing philosophies [trend following versus mean reversion]…then every eight years you [mean reversion] are out of business because when it doesn’t revert to the mean, your philosophy loses…mean reversion is somewhat uninformed."

I think Joe may have hit the nail on the head, but I'll quote what I was about to say:

"I can imagine a situation where one is short volatility because Sharpe has been maximised (optimised) by exposure to high-beta stocks that happened to do well during a bull market, which then blow up when the market breaks down. If you are thinking of a different situation, please say it."

Sally, let us suggest for a minute that you have been given the empirical data from Quantopian: What are you expecting to see and what would be your conclusion? I assume that (absolute) minute-level returns will correlate with no metric other than volatility (why would it?). Daily returns I assume you are expecting some sort of negative correlation between it and Sharpe.

On another note, I can't think there would be many firms doing algorithmic trading prior to 1990. Where were you? It seems like the more likely setup would be Fortran and a telephone.

What has been will be again, what has been done will be done again; there is nothing new under the sun. Ecclesiastes

I have been very fortunate and have worked with some of the greatest minds and most successful hedge fund and money managers. I don't want to reveal too much about myself, I use my middle and maiden names to stay somewhat safe, but...

The first hedge fund started in 1949.

I read this 1967 book when I started in the industry. It shows how to create a warrant arbitrage algorithmic trading system and is an industry classic. Everyone should read it.

In the 1980s, a Columbia University professor of computer science invented the concept of pairs trading, launching a hedge fund in the late 1980s.

In October of 1987, I worked on the trading desk of one of the largest fund companies. We used all manner of computer driven trade algorithms. VWAP trading and ITG were 1980s innovations. I met Benoit Mandelbrot and worked for a time with a guy from Putnam on using fractal distributions in place of normal or log-normal distributions to model extreme market events in the late 1980s and for trading commodities.

I worked with a neural networks guy for stock selection in the mid 1990s.

Enough said for now and I hoped I answered your question about algorithmic trading prior to the launch of python. There is nothing new under the sun.

Sally, let us suggest for a minute that you have been given the empirical data from Quantopian: What are you expecting to see and what would be your conclusion?

I have no idea what conclusion I could reach before seeing the data. My experience tells me that the pattern of returns tells a great deal about what happens in the market, maybe everything. No more, no less.

My takeaway is not that there would be a negative correlation between daily returns and Sharpe, but that algorithms of similar Sharpe ratios would likely show similar performance under certain market conditions. This takes the "hedge" out of "hedge fund".

It isn't so much that Sharpe should be discounted, but that algorithms with high overall returns and poor Sharpe ratios (especially during market conditions when high-Sharpe algos are doing poorly) should not only be reconsidered for Quantopian's hedge fund, but encouraged.

Selling vol is the typical negative skewness trade, which makes small amounts of money most of the time but has periodic catastrophic losses (when markets seize up -- so the worst time to take a loss). It's a classic pennies in front of streamroller strategy which can be profitable if you've got the massive capital requirements and risk tolerance to weather the hits but is in no way a conservative strategy.

So if you could optimize true future Sharpe, it wouldn't be a terrible metric. However, if what you're doing is simply measuring historical Sharpe, especially over recent history, such negative-skew trades tend to look very well so long as long as your lookback period doesn't contain a crisis for the particular factor you're taking on. 'When Genius Failed' is a great book about LTCM, a fund which adopted a wide range of negative-skew strategies, had an amazing few years and then collapsed in a most spectacular fashion.

http://www.globalsigmagroup.com/Past_Performance/Past.html

Look at Dr. Rao's performance shorting Vol.