Below is a graph of the ratio of YLDEARNYEAR (SI variant) vs ^N1T (QQQ equivalent). It is a screen I used for several years but stopped using in 2019 due to poor performance. It appears to be now performing better than QQQ (positive slope of graph).
How do others determine when to stop using a screen??? Or start using a screen???
That is part of my dilemma. Over what period should a screen work to be considered a screen to use. Every day? Month? Year/ 5 Year period, life of backtest?
The backtest of this screen since 1987 is good. From 2014 to 2019, poor, since 2019 good.
You know Craig that is a very good question. Think about it, even the great Warren Buffett isnât in the green all the time, but he is still following his method. Day in and day out he doesnât change but that doesnât mean he is always making money. So that is where I surmise you might be. Not always in the green but still following your screen until you donât. So that is a perplexing problem, just when you change your screen comes back into vogue. Itâs sort of the market timing problem.
Interesting. Iâd suggest looking at the âfactorsâ in the screen definition and try to determine which one(s) may be picking up renewed usage by the institutions, from the general market buy themes.
A few years ago I theorized about a few reasons for YEYs retreat to mediocrity. The post was Message: MI Board Search mid=34696372 in a thread titled âold screen testingâ.
Particular to YEY I theorized that
in a ZIRP / TINA world, dividend yields had been declining and higher yield stocks were not a worthwhile factor to attract demand
the screen had become dominated by MLPs, which were high yield but not high growth and - as mungo has pointed out - profits dependent on a commodity price.
The âfactorâ world has exploded and money moves much, much faster since most of these screens were defined with their traditional fundamentals-based / momentum combination.
For YEY, since the covid crash of 3 years ago, everything has been doing well, but we could imagine (and its been reported that) quality higher-yielding stocks have attracted more buying as interest rates have risen - bottomed before the Fed went to hyperspeed, perhaps. Saas and hypergrowth and the FANGs were all the rage up until '19 and the end of 20⌠not so since then. The moneyâs gotta go somewhere.
I really, really miss Bill2mâs tracking of every screenâs week to week performance so that it would be easier to figure out âwhatâs working latelyâ.
If the past data you are using to backtest is representative of the future, then your screen should work. If that past data is no longer representative of todayâs investing world (which could be due to any number of many variables), then maybe your screen might not work.
In general, we can probably agree that the investing world is always changing to some degree. When I read about someone analyzing data back to 1920s to look at some pattern (which is the same exercise as a backtest), I always wonder how relevant that old market data is to today and tomorrow.
What I am saying is similar to what flying circus says, except FC is actually providing an explanation for how the historical data might differ from today.
I compare the screen to a benchmark, usually SP1500 equal weight. QQQ is mostly just the technology sector, and so might not be the best comparison. The OP chart doesnât say if YEY did well or QQQ did poorly from 2000 to 2010. Looking at rolling 5-year returns compared to SPY, QQQ underperformed before 2008, outperformed between 2008 and 2018, and had amazing returns since 2018. portfoliovisualizer
The OP chart has an interesting feature. The period [1990 to 2001] looks similar to [2010 to 2021]. Probably just tea-leaf reading. I made an overlay of these 2 periods:
The evaluation period will vary because of individual risk and portfolio differences. I use 3 timeframes: long, medium, and short. My first requirement is that the longest backtest Sharpe beats the benchmark. I then look at results over a recent full market cycle, for example March 2009 to October 2022, and also consider the most recent 1-year results.
I have added the performance of YEY and ^N1T to the graph shown below. The reason for ^N1T (QQQ equivalent in GTR1) comparison is that QQQ is what I purchase with money not invested in a screen.
For the 5 year period, 2014 to 2018, YEY averaged 7%, while QQQ did much better. 2021 and 2022 were good years for YEY.
Looking at the last 5 years, YEY has performed better, but only during the correction
First a question. What is the vertical scale? Did the screenâs aggregate performance out run the index by more than 100 to 1 over the whole test period? Or does it mean something else?
Looking at the chart, it seems that the screen also matched the index roughly from 1987 to 2000.
I would say that if a screen only matches the index during some periods, but rarely underperforms it, and outperforms during other multiyear periods, then itâs a great screen.
There is no good way to decide when to abandon a screen in favor of an index and when to come back, IMO.
The vertical scale is portfolio value. This type of chart is sometimes called an Equity Curve. On a semilog scale, the slope of the Equity Curve is CAGR. Long term YEY CAGR is much higher than the market, but YEY was flat between 2014 and 2021.
YEY is datamined nonsense, and will match market returns. Everything is random. Holding only a few stocks will result in noisy outcomes, with some portfolios doing well and some failing.
The market has changed and is more efficient now. Bargains are difficult to find. Some screens will work and others will fail. YEY has X% chance of higher CAGR.
YEY will revert to the mean, and have lower CAGR. Last yearâs hot stocks are poor investments.
YEY will return to glory with similar CAGR as the long term backtest. The screen has found a niche of the market that gives higher returns, perhaps because of higher risk or maybe because of some sort of behavioral bias in the market.
Thank you for the comments and additional data. Interesting to me that YEY had a lower SAWR going from bottom to bottom. More work, more anxiety and lower performance.
So I will continue to ponder whether to renew trading the screen.
Again: review the past picks - what businesses is YEY (or any screen of interest) do the screen factors drive it to be investing in? Whatâs the commonality? Is it a common sector, industry or just high relative strength (which especially in 2020 meant SaaS tech).
How does the performance of YEY-SI look - similar or vastly different? Does this finding hold-up in a similar but different universe or is this driven by the underlying VL Timeliness performance?
Looking at the graph which broadly covers two-three pretty large bull markets & the respective bear markets it seems that YEY outperformed in the following periodsâŚ
1990 - late 1994
2000 - 2006
2010 - 2014
Late 2020 - Early 2022
All of which could probably be termed early - mid bull market (from memory)? YEY seems to have notably under performed in the later bull market? Is there something about those periods we can identify - thatâs related to the definition of YEY for robustness perhaps - that predicts whether YEY will outperform or not?
That is mechanically choose whether to use the screen or not - a switcher strategy - so that the portfolio continues to perform in each period. Perhaps the Market Capitalization / GDP or itâs first / second derivative (i.e. the rate of change / acceleration of the measure) or similar?
As a general note though: Iâm reluctant to over-complicate things though. Iâd much rather take a simpler-but-less-fragile screen. Thatâs one of the reasons I personally stopped using YEY in favor of other screen blends.
Thatâs what I thought, but I had to ask to be sure, because the results are mind blowing.
It means that $1000 invested in QQQ in 1987 became about $100K. And $1000 in YEY over the same period would be about $30 million now.
I really, really miss Bill2mâs tracking of every screenâs week to week performance so that it would be easier to figure out âwhatâs working latelyâ.
FC
You can use lohillâs GTR1 Helper to track and download screen performance or just automate it yourself.
For several years I kept up a spread sheet with monthly returns for 26 MI screens. From this I built a screen selection backtester with a combination of momentum, std deviations and longer term and recent outperformance. Also chose the number of screens and positions to hold. Used an optimizer the pick the principal components and some hysteresis to prevent switching screens too often.
Optimum from back then was picking from screens that had a long term avg return / long term std dev with a little more emphasis on return than std. Then using a 6-month momentum with slightly more emphasis on recent.
It sometimes only picked a screen for 2 or 3 months but more often would stay with the same screen for 1 to 2 years. In fact, the majority of screens were never picked.
Overall this did slightly outperform any single screen but overall I didnât significantly outperform the market.
Here is a version using only SIP data based on a YrErnYr post by Platykurtic around 2015 which I converted to use in my screener back in 2015. My backtester says it has performed very well in last 10 years. Which of course means I havenât been using it!
Note: The sectors and industry need to be updated.
Monthly:
and ci.EXCHANGE <> âOâ
and psd.price > 5
and CI.IND_3_DIG <> â721â
and CI.IND_3_DIG <> â933â
and (ci.company Not LIKE â%%L.P.%%â and ci.company Not LIKE BINARY â%% LP%%â and ci.company Not LIKE BINARY â%%LLC%%â and ci.company Not LIKE â%%partners%%â)
and MLT.PE > 0
and MLT.YIELD > 0
and PSD.PRICE*PSD.AVM_03M/21 > 1000
and psd.PRICE > (0.8 * psd.PRICEH_52W)
and MLT.YIELD/MLT.PE > 0.5
and ISA.DPS_12M > 0
order by (psd.PRCHG_52W + (100 + psd.PRCHG_52W) * (isa.DPS_12M / psd.PRICE)) DESC
Top 7 HoldTillDrop 10
The code only works for my Python based backtester/screener. However each step compares a SIP field to a pass criteria. The field prefixes are SIP field categories (psd = price and share statistics, mlt = multiples, isa = income statement annual) which you can ignore. I download SIP data from AAII every month since 1997 and weekly since 2010 and put it into a MySQL data base. This is a subset of the MySQL query that either runs a screen or retries data for a backtest. Not as full featured as GTR1 but adequate for my needs.