Quantifiable Edges Recent PostsView All
I have begun sharing one of the Quantifiable Edges Seasonality Calendars each month. Last month, we looked at long-term treasuries . This month the calendar that caught my eye was the NASDAQ Composite Index.
The Quantifiable Edges Seasonality Calendar uses multiple systems to measure historical performance on similar days to those on the upcoming calendar. The systems look at filters like time of week, month, year and so forth. Over the long run, staying out of the market on days that do not appear in green, would have been beneficial. To appear in green the date needs to show a historical Win% of 50% or more and a profit factor of 1.0 or more.
Obviously the NASDAQ stood out because its Calendar is almost all green. But not only do we see mostly green this month, I’ll also note that almost every day up until April 23rd we see numbers above the “baseline”. The baseline is simply NASDAQ stats over about the last 10 years. So the next 3 weeks we see that NASDAQ seasonality is both positive, and mostly better than average. Traders may want to keep this in mind along with other factors they consider as they establish their market outlook over the next few weeks.
Want research like this delivered directly to your inbox on a timely basis? Sign up for the Quantifiable Edges Email List.
In this writeup I am going to explain how to build a simple calendar model that uses machine learning to determine the days with the best odds to trade. The I will show how the calendar model does versus “Buy and Hold” over a long time period. The approach here is similar to how the Quantifiable Edges Seasonality Calendars are produced each month. Of course they are a bit more complex, and they are also enhanced by the fact that rather than using just one model, they use an ensemble of models to generate the statistics. But as you will see, even a very simple calendar model can produce surprisingly effective results.
Below are the steps to build the model.
I’ll go through these briefly one-by-one.
Data is sorted by calendar day of month (1 through 31) – Here we are simply establishing groupings. In this model, the grouping are extremely simple. What day of the month is it? You could also use trade day of month, or normalize the number of days, or segment many other ways. But the idea is to determine whether certain groupings provide an edge vs others. Here we are just doing “day of month”.Performance is measured over a rolling 10-year period – To see what groups provide an edge, you need to measure the performance of each group over a specified time period. In this case, I chose 10 years. With 31 groups and 252ish trading days per year, this means each group will have about 81-82 instances, though a little less for the 29th – 31st . Performance can be measured multiple ways. For instance, Win %, Avg Gain, Profit Factor, or by any other performance measure you deem important. The time that is measured can be whatever you determine appropriate. I used 10 years as a nice, round number that is long enough to typically include both bull and bear market phases. You can use much longer or much shorter.Stats calculated and tracked for each period – Once ten years (or whatever length you choose) of data is available, the performance stats can be generated. They should then be rolled forward continuously. This will allow the machine to adapt to changing market conditions, and for the old data to roll off, no longer including it in forward decisions.Stats for the upcoming day are used to determine whether to be in or out of the market on that day – So if tomorrow is the 5th of the month, the model will look back at performance of the 5th of the month over the last 10 years to determine whether to be long or flat at the close today. In the results I am going to share, I required a Win % of at least 50% and a Profit Factor of at least 1.0. So we are looking for the stats to be neutral or positive in order to have a long position. Otherwise, we get flat.This is a machine-learning approach. We don’t tell the model the best days. The model finds them itself – Again – nowhere in the code do we specify that we view the 1st or the 5th or any other day as bullish. Bullish/neutral/bearish are evaluated on a rolling basis by the code.
These results are fairly remarkable. The model is only exposed to the market about 52% of the time, and yet the annual return beats the market by 1.35% per year. When it is out of the market, it has earned interest at the rate equal to the 30-day Fed Funds rate. (I used that rate since 1) it is generally the lowest published rate, and 2) I had data back to 1957. Prior to 1957, 0% interest was earned by the model.) Over a long period of time, the 8.63% CAGR vs the 7.28% CAGR makes a very big difference. Also interesting about this is the fact that it is 100% in or 100% out. There is no ability to leverage in this model. So it cannot outperform when the market is rising. It can only outperform by side-stepping unfavorable days. And since it did outperform, you would think it did a decent job of reducing drawdowns. Below is the drawdown chart.
The model is represented by the red line and the SPX by the gold line. Interestingly, in almost every major drawdown over the 81 years, the model managed to avoid a portion of the drawdown. And that is how the performance difference became so large over time.
As I stated earlier, this is similar to the approach used by the QE Seasonality Calendar models. I also thought it was a decent example of machine modeling. (Note: This approach can be used with any kind of data. It does not need to be calendar-based.) I also thought people would find it interesting how a simple day-of-month filter could be so effective over such a long time period. Day-of-month seasonality IS a thing. And based on this, it appears worth paying attention to. I hope you found the above exercise thought provoking.
If you’d like to learn more about the QE Seasonality Calendars, feel free to check out the Intro Video or the Quantifying Seasonality webinar on the QE Youtube page . You can also take a free trial of Quantifiable Edges , where you can download white papers and more information in the Seasonality section of the subscriber area.
Want research like this delivered directly to your inbox on a timely basis? Sign up for the Quantifiable Edges Email List.
Quantifiable Edges currently has 2 premium models whose signals are available on a subscription basis for individual traders. They are the ETF Momentum Swing and the Duration Rotation models. As of 3/31/21 we will be closing those models to new subscribers. After that, they will only be available through Capital Advisors 360, LLC separately managed accounts program. If you would like to license either the ETF Momentum Swing, or the Duration Rotation models, you will need to do so by 3/31/21. Information and pricing can be found using the links below:
https://quantifiableedges.com/subscribers/signup/ETFmomentumSwing
https://quantifiableedges.com/subscribers/signup/duration
If you would like more information on the Capital Advisors 360, LLC separately managed account program to determine whether it may be suitable for you, then you can contact Rob Hanna at robh@capitaladvisors360.com
Over the last few months bond investors have been hit with the worst combination of rate moves. Intermediate and long-duration bonds have been rising rapidly while the very short-term rates have actually continued to decline. This can be seen in the charts below, which show the 30-yr Treasury Rate, 10-yr Treasury Rate, and the 13-Week Treasury Rate.
The rise in long rates has led to TLT (iShares 20-yr Treasury Bond ETF) losing over 20% from its 8/4/2020 closing high – that includes dividends paid over this time. A decline that large can be tough to make back when the current yield is just 2.1%.
But the drop in rates on the short end has also been painful. Many short-term bond ETFs are struggling to keep their yield above their expense ratio. And the only way they are managing to do so is with a big chunk of their bond holdings in BBB or lower rated securities. Of course yields below expense ratios over an extended period means a negative return is likely in those assets. Below is a list of some of the more popular short-term bond ETFs along with their stated 30-day SEC yield and their expense ratios.
The Fed is not likely to remove pressure on short-term rates anytime soon. But right now it means bond investors are losing on the long end, and are even struggling to breakeven on the short end.
Want research like this delivered directly to your inbox on a timely basis? Sign up for the Quantifiable Edges Email List.
Monday was the 7th day in a row that the DJI closed higher. This triggered a study from the Quantifinder that looked at performance after 7-day win streaks in the Dow Industrials since 1980. I’ve updated the stats table below.
There is not much of an edge over the 1st few days. But once you get out a little further, the stats appear solidly bullish. The 16-19 day returns show a very high win %. Momentum tends to carry, and this study is just a simple example of that concept. Traders may want to keep it in mind as one factor when determining their intermediate-term market bias.
Want research like this delivered directly to your inbox on a timely basis? Sign up for the Quantifiable Edges Email List.