I recently received the following question from ITA member Jim Hotvedt:
“How sensitive should rankings be using the BHS model with BL/SH set to “Yes”? I get the exact same rankings (top 4) with the Buy Low/Sell High Period set to 10, 20, 30, and 60.
My other settings on the Menu tab are LR1=60, LR2=100, Projection Period=100, Convolution Period = 10, and the HAs set to 3 and 5. The same thing happened when I changed LR1 and LR2 to 22 and 65, respectively.”
Since I thought that other ITA members might be interested in this excellent question I thought that I would write this dedicated response in an attempt to answer the question. As usual, in this game, the answer isn’t easy/straight-forward – so I’ll try to answer the question as best I can.
The first thing to remember is that no system works perfectly under all market conditions – so, what works best in one time period may not work best in other time periods. This is the challenge of investment/trading.
I will try to illustrate my response with some illustrations. These may not be the best examples to try to explain my response – but they are current (and not cherry-picked).
First, let’s take a look at the LRPC graph for VTI using the “default” 90-day look-back period for the Convolution line:
If we look for the LRPC crossover signals (red and green arrows with black outline) we see three signals in this recent 12-month period (that includes the March “melt-down”). When we look closely, we can see that the “Sell” signal was indicated a little late near the end of March. The re-entry “Buy” signal also came a little late in June. Prior to that we only saw one “Buy” signal (in September) that might have kept us in the market (with no trading) for ~6 months. The rapidity of the (~20%) pullback in March was a little too fast for the default 90-day look-back of the Convolution line – but, remember that this was chosen based on longer-term back-tests, covering many market environments, and was found to generate the most consistent long-term performance/returns.
Now, if we change the Convolution look-back period to 10-days rather than 90-days we see the following graph:
where we see a lot more action in the (red) Convolution line. However, we note that the recent Sell/Buy signals were generated much earlier that in the first graph – early March and late May respectively.
So. Why would we not just set the default look-back to 10-days instead of 90-days? Well, take a look at the “jittery” action in the Convolution line in the September-November period where we see two “Buy” signals and one “Sell” signal – this kind of action usually results in “whipsaw” trades and can severely degrade longer term performance. This isn’t the greatest example of “whipsaw” action, but is, hopefully, good enough to illustrate the point. Remember that the 90-day look-back in the first figure (just) managed to avoid this action.
While we are looking at these figures, note where the dashed green line crosses above/below the zero line. This line plots the slope of the projection line (with a 10-day look-back). As can be seen in the second figure this zero slope crossover occurs at essentially the same point as the LRPC crossover when both look-back periods are set to 10. Thus, using the BHS system with BL/SH set to “Yes” includes this signal in the “score” rankings rather than the default 90-day LRPC signal (with BL/SH set to “No”) – although the LRPC Convolution look-back could be set to 10 (and BL/SH set to “No”) to generate the same result as BL/SH = “Yes”. ‘Hope that isn’t too confusing – it needs thinking about.
To summarize, let’s take a look at the settings in the MENU sheet:
The blue arrows show where the parameters can be changed for the Convolution look-back (left-hand side) and the BL/SH look-back (right-hand side) periods respectively.
So, in response to Jim’s specific questions, the sensitivity of rankings to the parameter settings will depend on market conditions. When markets are “quiet”, using short look-back periods will probably result in unwanted “whipsaws” but, during (or anticipating) larger/faster price movement they are likely to be better. At the present time, looking at the LRPC charts, we probably shouldn’t expect to see too much difference in recommendations if we change the parameters.
Similar arguments go for LR1 and LR2 look-backs and for HA settings.
‘Hope this helps in understanding some of the more subtle aspects of the Kipling Workbook.