More real work done by Herb Haynes ?

In Parts 5 and 6 of this series of posts, we looked at the relationship between the Projection look-back period (L1) and the Convolution look-back period (C1) – keeping all other parameters constant. In Part 3 we described the Convolution parameter and the fact that it comes in two parts, a look-back period (C1), for purposes of calculating the slope and intercept of the linear regression line, and an offset period (O1) to determine at what point on the line the Convolution value will be measured.

In this post we look at the relationship between C1 and O1 in terms of their impacts on system performance. Again, in order to determine these relationships, we must keep all other system parameters constant. Thus, all parameter values used in previous tests remain unchanged – but, specifically, of the parameters we have already looked at, we keep the ** Projection look-back period (L1) fixed at 100 days** and we keep the

**.**

*SHY filter turned off*As before, let’s look first at total returns as we vary C1 between 20 and 100 days and O1 between -20 and -100 days:

As we would expect, the value in the highlighted cell (L1=100, C1=90, O1=-50) is the same (234.2%) as the value reported in Part 6 of the study. Here, we get confirmation that -50 is a “robust” value for the convolution offset parameter – in fact, combined with L1=100 and C1=50, results in “optimal” performance (245.1% maximum total returns) – but, we’ll stick with the highlighted cell for focus.

Obviously, CAGR values reflect the total return rankings in terms of performance:

Now, let’s take a look at portfolio volatility:

Again, we see that the highlighted cell falls in a compromise “moderate” position – not in the highest percentile yet not in the lowest percentile ranges.

So, let’s look at the return/risk balance through the Sharpe Ratio:

…. Looking pretty “robust”.

As in previous posts we can move on to look at draw-downs:

Our highlighted cell remains in the top percentile green areas with DD ~15% – but there are lower DDs towards the bottom right hand corner of the map.

So, let’s balance returns and maximum DD by looking at the MAR ratio:

Here we see the penalty paid if we want smaller DDs – lower total returns/CAGR (top 2 figures).

The following HM shows the calculated number of trades in the above tests:

Ok – so we “cheated” a little when we fixed the value of O1=-50 in Parts 5 and 6 – but, had we chosen something else (e.g., O1=-20), we would have arrived at the same point at this stage of our analysis.

These tests just confirm/justify our choice of -50 for the offset parameter.

All these parameters (L1, P1, C1 and O1) can be assigned/adjusted in the MENU sheet of the LRPC workbook. Although P1 is a parameter that could be changed, we keep this constant at 20 days since this is logical and appropriate for a monthly review schedule.

Now that we have taken a look at all the Projection and Convolution (PC) parameters we can select values that we might favor and take a look at the impact of changing the number of assets to be included in the portfolio. This will be the subject of our next post. Remember, all our tests to date assume no limit to the maximum number of assets that might qualify for inclusion in the portfolio (10 for the Rutherford list) – subject to the positive P-C value requirement.

As usual, downloadable PDF file at https://www.dropbox.com/sh/ix9tocyjyf4ycl3/AACoRWSYilTa4Eu6sUi60GI3a?dl=0

Herb and David

Robert Ryan says

David, Herb,

Great experiments, results and things to think about. Not to be too picky but which “Rutherford” is used in the back-tests? In July you replaced TIP with AGG.

” The only other minor change I want to point out here is that I’ve replaced TIP with AGG in the Rutherford asset list. TIP is primarily an inflation hedge and with inflation not showing signs of moving significantly higher in the short term and also with Gold in the asset list (as an inflation hedge) I’ll replace TIP with AGG (slightly better dividends) going forward. I’ll keep TIP in the back of my mind to move back into the list should inflation show signs of rearing it’s head.”

Yet the correlation between TIP and AGG is only 0.6 or so over the last 10 years and inflation just might be rearing its head. How would the test be if TIP was used in the Rutherford 10 rather than AGG? Sure don’t want to overly tune any model, but isn’t there a risk of prejudging the future to accommodate the present? If I do want to have an asset class which has low correlation with the other 9 in the Rutherford 10 and does pay a dividend (and has “low” SD) why not add TIP to replace GLD rather than AGG replacing TIP? The results would be? The same? Hope so.

Just curious,

Robert

HedgeHunter says

Robert,

Yes, these tests included AGG rather than TIP in the Rutherford 10.

I don’t have any problems with your suggestion to replace GLD with TIP – they (TIP and GLD) are both inflation hedges – but, then, I have access to more back-test results than you have seen, so I might be accused of curve-fitting. We haven’t run any tests with TIP – but my gut tells me that you might be right – and we might see better performance. Remaining neutral/unbiased over a period where inflation hasn’t been a factor is difficult – I think we would want one “inflation hedge” in the portfolio – TIP or GLD – no matter what the market environment. If I were to take GLD out, rather than TIP, I might be inclined to replace it with IEF rather than AGG – but now we’re getting a bit picky. There is obviously a difference in volatility between GLD and a bond fund – so a longer term bond (IEF vs AGG) may not be a bad decision.- but now we’re really down in the weeds and susceptible to accusations of curve fitting.

Our intent has been to try to use a “benchmark” portfolio for system comparison rather than to “optimize” the Rutherford.

But great comments/observations…

David

Arthur Cummins says

These are very informative post. I’m looking at this and the previous two posts and comparing it to potential data entry points on the menu page. It seems that the terminology is not consistent to me. The bottom of this page says we can L1, P1, C1, and O1 on the menu page. Is L1 the Linear Regression Projection Period? C1 the Linear Regression Convolution period? P1 the Extension Period? and O1 the Offset Period? Therefore, the Kipling LRPC 1.1.1 menu pages shows these as a default value of L1 – 100, P1 – 20, C1 – 90, and O1 as -50. Is that correct and is that consistent with the sweet spots in these blogs?

HedgeHunter says

Arthur,

Yes your interpretations are correct on all points 🙂

David