Skip to content

Is the Day of the Week Phenomenon Tradeable? It Was For A Long Time But Not Any More!

September 29, 2009

So to continue with our little time machine experiment, i fed the Day of Week data into the learning algorithm to see whether it should be traded or not. Result? I got a bright red light–which means that Day of Week effect is completely un-tradeable-(ie not even worth going short the strategy). I tracked back the different “lights” and discovered that  1998 was the last year  it was considered worth trading.  Without digging too deep into methodology,  the Adaptive Time Machine takes “de-trended” data–which means that it looks at daily returns (or whatever interval is being investigated) net of the average yearly return so that it does not mistake trend effects for the specific strategy that we are testing. In addition, a separate learning algorithm detects changes in the equity curve of the strategy itself—similar to what a human would do: 1) seeing whether or not performance is accelerating/decelerating 2) seeing whether or not performance is becoming erratic. Of course, this filter makes judicious use of statistics, but the concept is the same.

As you can see in the data presented below, the Day of Week strategy was in fact tradeable for a long time prior to the last 3000 bars. However when we take a look at various confidence filters, there appears to be little or no relationship in the last 3000 bars between confidence and average daily return. If anything, performance has actually been negative when confidence was high that a given day was positive—-partly this may have something to do with the mean-reversion effect that set in this decade and the new mean reversion “weekend effect”.  For a great post and background reading check out  http://marketsci.wordpress.com/2008/10/07/weekend-vs-weekday-follow-through-does-friday%e2%80%99s-sentiment-carry-over-to-monday/ Although it is clear that nothing inherent in the ranking of a given day of the week includes any information about the previous day’s return any more than the other days. However, if that was the predominant reason, we might see a clear reversal or a greater linearity of effect to the downside. Obviously a great deal of noise has been introduced in recent years into daily returns, and part of this in my opinion is the creeping onset of increased market efficiency. The second set of charts show the use of a ranking methodology to identify day of week effects and also to create trading strategies. De-trended data was also used for these examples. Ranking each day of the week using a statistic called the DVR (Sharpe ratio x equity curve R-squared), there was a distinct–although not perfect–relationship between the rank and the average daily return out of sample prior to 1997. It is now a little clearer to see the randomness of the day of week effect in the last 3000 bars both in absolute and risk-adjusted returns. Even clearer is comparison of the equity curves of two long/short day of week portfolios both prior to and after 1997. So to answer at least one question: Is the Day of the Week  phenomenon tradeable—not any more! So, the lesson is—always keep an open mind, but also make sure to dig deeper into strategy analysis and it is critical to adapt and revisit your analysis as time goes on.

DoWconfperf

123

DoWrankperf

123

130-30DoW60-97130-30DoW3000

Advertisements
4 Comments leave one →
  1. September 30, 2009 9:58 am

    David, I know the concept of detrending is very simple, yet I still fail to comprehend exactly how the data is detrended, and even more, WHY detrend the data? Perhaps you could flesh that out in the comments, or maybe even a post, when/if you have time. It might be helpful for folks like myself that are on the cusp of understanding all this but yet still have a ways to go.
    Thanks!

    • david varadi permalink*
      September 30, 2009 10:58 am

      hi wood, it sounds more complicated than it is—all you are doing is say taking the 252 day average daily price return (or 200 day) and subtracting it from the daily return to take away the trend component. The reason is you don’t want an upward or downward bias to obscure a given effect. Very simple. This concept can be extended to a lot of different methods depending on what you are trying to isolate. Of course fancier methods can be used, but this method of de-trending works well.

      cheers
      dv

  2. October 1, 2009 8:00 am

    Thanks David. That is what I needed.

Trackbacks

  1. Can Seasonality Be Traded Profitably? A Time Machine Test (Part 1 of 4) « CSS Analytics

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: