Here's the first algorithm, which uses the GARCH(1,1) to signal shorting the VXX. Clone it and try it out with some other models.

And the second, which uses Quandl Data find the settle price of VX1 futures, then uses that as a long/short signal.
Clone this one and implement some other interesting models!

As an aside, the comment on line 95 of this algorithm is no longer relevant. It also appears on line 80 of the first.
# Use the look-ahead data as a defacto volatility prediction

There is no imported data in the first algorithm, and the quandl data is time-shifted in the second. (On line 40)
f = df.tshift(1, freq='b')

It took me a couple of reads to really get the relationships you/ernie are describing, but I think I have a simpler way of explaining:

When today has a negative return on SPY, it's likely that implied volatility has increased today. People were scared into buying more insurance. Generally, implied volatility at an elevated state like this tends to fall the next day. Thus a negative return on SPY is likely to be followed by a reduction in implied volatility the next day.

A positive return on spy tends to be good news, and no one feels the need to buy more insurance, so implied volatility also falls the next day.

Generally speaking, implied volatility spikes quickly but falls slowly, hence on both good days and bad days for the market it can still fall on average.

The situation you describe with a very negative return on SPY might not be that statistically significant. I haven't checked.

My trading strategy takes the following route:

First, a good predictor for future realised volatility is historic realised volatility. Yesterday's weather is a pretty good predictor of today's weather.

If implied volatility is higher than this prediction, then it's in an "elevated state" and is likely to come down, so you short VXX. Otherwise long VXX.

Sorry, I should point out that implied volatility is usually a few percent higher on the long run than realised volatility. This means my strategy has an inherent bias to shorting VXX (i.e. selling volatility insurance) which is good, as it then benefits from the usual contango in VX (aka volatility risk premium).

"Shorting the VXX" is a "sure" way to make money and you don't need an indicator to do it - just keep shorting and stay put and wait.

The only small caveat is that this is not possible to do - Quantopian lets you "short" it, but your broker will not, since it is HTB or NTB (hard to borrow or none to borrow).

There is nothing new under the sun when it comes to short vol strategies.

You can go long XIV and rebalance every day, which is pretty close to shorting VXX. There is the risk that fund blows up if VIX spikes sufficiently, but at least they manage that risk intra-day. Much safer than shorting VXX yourself on Q!

Buy and hold XIV is one strategy I've seen to harvest the VRP you've mentioned. It does suffer from very deep DD though, and I think even a simple trading system like the one I use can improve this performance.

Simple strategy attached.

In Ernie's presentation he suggests this strategy:

RV(t+1) is the GARCH-predicted realized volatility for next period. If this is greater than IV(t) = VIX(t), then buy VXX and vice versa.

I've replaced the GARCH model for realized volatility with just the standard deviation of the last 5 days. Might be fun to see which predictor is more accurate!

Note, the huge IV spike looks like an error in the Yahoo VIX data.

@Lotanna I think you have an error in your code. You are using the price of SPY instead of the log return?

Hey, Is there any way through which I can reduce volatility of the generated returns?

@Hemant

Is there any way through which I can reduce volatility of the generated returns?

Certainly, just use less volatile instruments or reduce leverage.

Below is tear-sheet of the same strategy using less volatile instruments QQQ and IEF.

Hi Vladimir,

That was great help. Can you share the algorithm as it will provide more clarity in terms of your thought process.

@Hemant,

Can you share the algorithm ?..
By some, known only to Quantopian staff, reasons I do not have ability to attach my backtests to posts since April 2017.

@Vladimir: Do you have the ability to paste code into your posts?

@André.

I just replaced XIV by QQQ, VXX by IEF in last Burrito Dan code and run back-test from 2007-06-01