Hi all,
It’s been nearly two years since I launched the “Curve Fitting for Fun and Maybe Profit” Designer Model. Let’s see where we are.
THE IDEA
Curve fitting on Portfolio123 has been occurring, and I’ve come to recognize certain factors as being key components to the curve fitters’ palette – and I created a ranking system named just that, “Curve Fitters Palette,” based on those factors. Most of these are used because, although they make no sense at all from a financial standpoint, they have often showed well in simulations of the past. Others are legit for those who really want the styles they represent (mainly the price and size-based factors) and were included because they tend to be favored by curve fitters given the way conditions post 1999 allowed them to flourish (meaning their presence tended to boost simulations that started in 1999).
Saying the factors are no good, etc. is all well and good but the reality is that we live, invest and trade in a world ruled by supply and demand so financially speaking, it can be argued that “might makes right.” If enough people think a nonsensical ratio like Pr2SalesQ is worth using, does might-make-right turn it into something that ought to be considered? Put another way, using the language of the p123 virtual strategy design course, can such items be used to harness and potentially work with the ever-present and more-important-than-many-realize “noisy” component of stock prices? (Besides P123, who knows how many elsewhere are working with similar factors.)
I tested it by applying this ranking system (which is now publicly visible: [url=https://www.portfolio123.com/app/ranking-system/272364]https://www.portfolio123.com/app/ranking-system/272364[/url]), not to an entire universe but to a subset that has been screened in a modest way for basic liquidity and some simple rules using sentiment and quality (the latter not being an effort to grab the best but instead to weed out the worst).
THE OUTCOME
As you can see on the Designer Models page, this strategy closely tracked the Russell 2000 ETF since it went live on 9/11/15. That’s actually not bad given the handicap of an equal weighted portfolio versus a market cap weighted benchmark at a time when big caps have been in favor. So based on this back-of-the-envelope ranking system (I didn’t even bother to think about weights; I just made everything equal), it looks like there may be something to the notion that constructive work can be done using these factors as a way to harness market noise.
A simulation that replicates the Designer model is also publicly visible: https://www.portfolio123.com/port_summary.jsp?portid=1498864
I’ve long believed and argued that screening is critical to strategy building, that when we rank, we push the statistics too far when we ask the top 10 stocks to be better than the next 10, the next 10 to be better than the 10 after that, etc. I see screening (buy rules) as the equivalent of a pre-qualified prospect list purchased by a salesperson as an alternative to prospecting from a telephone directory. When serious screening rules are used, you limit application of a ranking system only to a subset of the universe that you believe has some potential to blend well with the ranking logic.
I decide to establish a second sim in which I took out the buy and sell rules that didn’t related to liquidity. I had to have something for selling, so I added Rank<90. That sim (which matches the live dates of the Designer Model version) is here: https://www.portfolio123.com/port_summary.jsp?portid=1498866
As you can see, it’s bad.
Finally, I set up a third version in which I even eliminated the liquidity rules. There are now no Buy rules at all and only one rank-based sell rule (Rank<90):
https://www.portfolio123.com/port_summary.jsp?portid=1498874
The third one is also pretty bad.
MY TAKEAWAY
I’ve seen what I wanted from the Curve Fitter Designer model and believe it’s time to put it into retirement, which I’ll do in the very near future. (I opened everything up here in case the few who follow it want to replicate it in their own accounts).
As to curve fitting, if we recognize these factors for what they are, it looks like there is a sound basis for their use. The key is to go in with eyes open and recognize how they might be helpful; as momentum-oriented factors that help one ride the N part of the P = V + N (price = value + noise) equation. And as momentum goes, these can be particularly interesting since most who work in this style stick to prices only. Harnessing something we might call “fundamental momentum” might lead to less crowded momentum trades.
One caveat though: Please take the Buy-rules seriously. That can help you a lot. Think about what kinds of stocks might be most amenable to momentum plays like this and apply the ranking system to those. Admittedly, the “good” version of the model was nothing to write home about, but it accomplished a lot using rules I spend less than a minute thinking about. It would be interesting to see what might result if more effort were dedicated to creating good sets of Buy and Sell rules.
