In a model is it a virtue or a vice to detail out factors?

Selecting one factor at random for example, Free Cash Flow Margin. Lets say i have rigorously back-tested and discovered that in order of effectiveness is; latest quarter, 3 year average and change in trailing twelve month from previous trailing twelve months. I have not done so for this factor, it may be that latest quarter has been priced into the market and may be in another order. The first may be notable in its effectiveness and the last may be close to worthless. I will go through the process of weighting the factors accordingly.

I am curious where everyone lies on the spectrum. I find catching myself apply this format to every factor in the same vein. I am wondering what everyone here thinks? I am rationalizing this to myself in the following manner:

Latest Quarter: A company may have notably improved/regressed in say its Free Cash Flow Margin and it may be always the most important aspect to consider.
3Y average: It may always be important to consider its long-term performance in a factor
Change (TTM from PTM): It may be important to consider in a change yearly

I have a sneaking suspicion i can streamline these elements, say change in this years Free Cash Flow Margin from last years in one or two factors. However, if its not hurting but just marginally better than one i might just do it for better comfort of mind.

Is this a virtue or a flaw?

I’m not sure I follow totally what is being described, but FCF for many companies can be pretty erratic, so I’d expect results based on latest quarter fcf likely benefit from being balanced against some annual or longer term FCF results. (fwiw, I use many variations of fcf over many timeframes). Short term changes in margins of all kinds are useful to me.

That said - if I recall, it is possible to build models that backtest very well just using very short term data and no long term data at all. I attribute it to variance in short-term expectations vs the market tending to discount longer term data better than recent short-term changes, but I don’t know - it’s just how my mind has tended to frame it after working with the database.

An exercise that I think I’ve tried, that might be fun to try, is to try building a model using only quarterly results. Then try to build a model using only annual results. Then maybe try a model using only 3yr results. It may give a feel for the strength of the different timeframes.

That exercise is pretty interesting, i will try that. Thank you.

FlynnTilbrook - I am a big advocate of factor/formula composites for situations like yours. Consider building a composite of the factors you mentioned and give each an equal weight if their advantage fluctuates from year to year. You can optimize the weights of each factor/timeframe, but be careful not to curve-fit. Sizeable composites (i.e., the number of factors) cover more bases than using one or two indicators, provides a reasonable average of the measures, and mitigates the potential for curve-fitting.

I use signals from composites of as many as 30-to-50 factors, providing broad risk mitigation that determines the appropriate level of market exposure for conditions. Those same composites can often be used as ranking systems to identify the highest-ranked stock or ETF with the least risk.

You might also consider a composite for different time frames for each of your key factors, in this case, FCF. This approach provides significantly improved robustness but may result in slightly lower Annual Returns. Many model builders want to get the highest possible AR and would discard this approach. However, by focusing on maximizing performance instead of minimizing risk, you decrease the probability of getting the highest possible return because you are perfectly predicting the past—rather than creating robust returns based on logical principles. It’s unlikely the market ahead will perfectly repeat the market behind, and you shoot yourself in the foot by trying to creating a model based on your highest performing sim.

This approach to risk mitigation significantly reduces drawdowns and results in equity curves like this live model that’s been running since 2004 (starting in mid-2007 would attain a CAGR of about 25% (the market has changed since 2008 because of constant and massive Fed liquidity infusions):

Our research shows that minimizing risk and drawdowns is far more important than reaching for the highest possible return, upon which most designers focus. After all, it’s not what you make—but what you keep! The chart above shows the S&P 500 using multiple composites focused on minimizing risk—with an average of only about two trades a year (7.25/month avg. hold time).

Chris

Thank you for your insights Chris. All your points have been very interesting and helpful.

Chris, what would you credit more for the minimal DD in your ETF model?
The fact that its an ETF, or your selection factors?

IMNSHO, it is the combination of using key risk-mitigating indicators with ETFs. I’ve been investing using ETFs since about 1995, but I also spent many years (1982-2017) investing in individual stocks, then in 2017 I dropped individual companies altogether. One big reason: ETFs eliminate the idiosyncratic risk that accompanies all individual-company stocks.

Your best possible analysis can pick what appears to be a steady, consistent, high-quality stock, and you might still see it plummet based on bad news or disappointing performance-related information. Individual stock prices can crash because of a disappointing earnings report, an FDA denial of a drug application, loss of a key executive—the reasons are endless…

To mitigate this risk, investors are forced to diversify with a large number of holdings—as many as 50-100 issues are required to reduce the potential for loss from an individual stock to -2% to -1%. However, that uncorrelated diversity usually muddles the portfolio’s performance to the point it’s not worth the hassle of monitoring all those positions.

With an ETF, you can feasibly have just one holding that spreads your risk over dozens, hundreds (e.g., SPY – S&P 500 ETF), or even thousands (e.g., IWM – Russell 2000 ETF) of individual companies. Also, you have virtually no risk of a diversified ETF plummeting to $0, as is the case with individual shares. The only possible way a diversified ETF (such as SPY) could drop to a worthless level would be for all 500 of those companies to simultaneously go under. A nuclear holocaust is the only way I can see this happening, whereas, with an individual company, bankruptcy can occur at any time.

I also find that an ETF-based portfolio is far more responsive to factors/formulas than individual stocks. In other words, I can more effectively leverage the information that’s available through P123 to build models with significantly improved results. So the combination of inherently diversified ETFs with composites of unique factors provides the steady, robust results you see in my chart (above).

Chris

Chris,
from 2004 to 2008 your model did not produce any alpha over benchmark SPY. Why is that?
Also why does the model start in July 2004.

Chris -

This is not really a “live model,” is it? It’s a simulation, right? When I read the words “live model that’s been running since 2004” I think of a model that’s actually been running live, in real time, out of sample, since 2004, when it was designed and launched (which is about thirteen years before you say you switched to ETFs). Maybe I’m being too persnickety about word choices, but I think it’s important to distinguish between simulated equity curves and real-time equity curves, especially in a public forum like this one. I’m sure this was just a slip of the tongue, but it’s best to be careful about these things.

Thanks,

  • Yuval

This is not an uncommon issue on P123. Adding a chart indicator that marks the start date of a live strategy would help.

This!

Because it was holding SPY. The average hold time is 231 days or nearly a year between transactions. SPY was in a bull market from 2004-2008.

I actually started running the model in 1998 as a SPY/Cash model. However, is 2004 I switched to SPY/TLT and started running it fresh from then.

It is a live model that I have been presenting publicly since 1998. I began investing in 1975 and took my first professional jobs as an analyst at Merrill Lynch and Drexel Burnham Lambert in the early 1980s. Private Equity and Hedge Funds followed. I guess all that makes me an ‘old fart.’

You are misreading what I wrote. I said, “I’ve been investing using ETFs since about 1995, but I also spent many years (1982-2017) investing in individual stocks, then in 2017 I dropped individual companies altogether.” …Meaning that there were 22 years of overlap in there before I abandoned individual stocks.

The model actually started in 1998, which would mean marking it before the chart begins. But with the 2004 revision using TLT instead of cash as a defensive asset, that is the live start date for this version.