It has very low beta.
Is it worthy?
Some times yes, but mostly no.

Hi Vadimir,
Thank you for your reply and making the backtest begin in 2007:). No disrespect. Maybe I was misunderstood, but I specified earlier that I was curious to know the point of view of the Q team & community regarding the type of strategy used, not if this particular algo was worthy of the fund. The algo for live trading is using more pairs of stocks and refreshing the stock list on a monthly basis. These are only some pairs that behaved well in the last 2 years, but nothing lasts forever.

For experimenting, condensed and sped up.

Groovy! Nice code optimization, garyha!:)

Thanks and hope it helps folks find improvements.
I need to make one correction, letting any be closed rather than just the first one:

def CloseSomePositions(context, data):  
    sales = 0  
    for p in context.pairs:  
        moves = MovesCalc(context, p)  
        offset = .003  
        if 1 - offset < moves < 1 + offset or moves == None:  
            order_target_value(p[0], 0)  
            order_target_value(p[1], 0)  
            sales = 1  
    if sales:  
        return 1  
    return 0  

Hi Adrian,

It's great to see you thinking about strategies for the fund. I cloned your post and ran it through the tearsheet to analyze the behavior. At first glance, the 2 year backtest from 2013-2015 looked interesting. However, as I ran longer backtests (5 yrs, 10 yrs) it appears the algo was overfit to the last 2 year period.

Attached is the tearsheet from the last 10 years. The returns could be improved, and the algo reaches a 50% drawdown. As you continue to test your ideas and create strategies, consider the bigger picture - Why does your algorithm perform well? What is the economic rationale? How does it perform in different market regimes? Feel free to share any other ideas for feedback! And if you find something you like, start paper trading it on Quantopian. Then you'll begin to collect out of sample data and track the returns.

Alisa

Hi Adrian,

Per your asking whether "this style of strategy is interesting to Q":
Yes, Definitely -- Everything you describe here sounds like a strong economic rationale to attempt to exploit with an algo. Specifically, as you mentioned: "correlated stocks (but not only) that tend to mean reverse when a major price gap (%) appears between them or that tend to keep wideing the gap / momentum, after a initial gap in price."

I think if you can extrapolate this idea into general logic for screening over a large universe of stocks, and possibly have, say, 10 - 20 pairs on at all times, the diversification benefit of many pairs should help mitigate risk in the overall portfolio.

Alisa,
What do you mean by market "regimes"? Also, how do you determine it was "overfit" for the past 2 years?

Thanks for your replies! You can find a similar shared strategy of mine, that trades only ETFs, here: https://www.quantopian.com/posts/october-prize-number-1-algo-disqualified-worthy-of-the-q-fund . Please share your code or trading strategy ideas on what can be improved to make it mode robust.
Adrian

@Adam, I'd be happy to explain. For the fund, we'll see how the algorithm performs in different market conditions, not only the performance in recent years. We'd like to see a backtest over 10 years (not an exact number) to see more data. As you're writing the algo, test how it performs in 2008, 2011 and other turbulent times. If the algorithm does well in the last 2 years, but loses money when running a 5,7, or 10 year backtest then it's very likely it was overfit to the recent data.

To prevent from overfitting, I encourage you to first think about your economic rationale. Why does this strategy work? What would cause this strategy to stop working? Then begin coding the details and testing over various periods. If you find something promising, enter it in the contest or start paper trading to see out of sample data. The more data it accumulates the better; paper trading is the true test.

The tearsheet is a very powerful tool to visualize the data. You'll see the length of drawdown periods, cumulative return trends, position concentration, and more. We held a webinar showing how to read the tearsheet plots and interpret the data; if you missed it check out this video: https://www.youtube.com/watch?v=-VmZAlBWUko

I think if you can extrapolate this idea into general logic for screening over a large universe of stocks, and possibly have, say, 10 - 20 pairs on at all times, the diversification benefit of many pairs should help mitigate risk in the overall portfolio.

This is a pre-appeal to give us enough computation time in the upcoming Modeling API to do the above! :) EDIT: woo, five minutes, thanks!

@Adrian,
I think Simon and Justin are giving good advice and pointing you in the direction of the new Pipeline API. We built it with the express goal of making it easier to extend algorithms like this to a larger number of securities. I hope it proves useful!