Turning a Research Paper into a Trading System (TRPITS), Part V
(Guest Commentary By Bill Rempel - February 7, 2008)
Dear Subscribers and Readers,
For those who had wanted to learn more about individual stocks, the art of stock selection, and model-based trading/investing, it is again time to see what one of our regular guest commentators, Bill Rempel, has to say. Bill is a prolific writing on the stock market and individual stocks and is the author of a very active market blog at: http://billrempel.com (“The Rempel Report”).
In this commentary, Bill is going to go into detail on constructing a feasible trading system catering to the retail investor – one that involves monthly turnover, trading in domestic exchange-listed stocks with over $100 million market capitalization. This commentary is a little bit more complicated and involve than most – but I assure you, this exercise is very rewarding for individuals who want to find an efficient way to automate at least part of their trading going forward! This commentary is a again fascinating read, but should only act as a starting point for one to do further research. Without further ado, following is a biography of Bill:
Bill Rempel (aka nodoodahs) is an active poster on the MarketThoughts forum as well as a few others around the web. Bill is a regular, monthly guest commentator on our website (see “Using Public Information to Find a Trading System Edge” for his last guest commentary). Bill graduated from Caddo Magnet High School (a high school for nerds) back in 1985 and proceeded to learn the hard way when he drank his way out of a scholarship to Tulane later that year. After a few years of sweating for a living, he decided to go back to school, and graduated from LSU-Shreveport in 1995 with a Bachelors in Mathematics - all the while working the overnight shift stocking shelves in a grocery store.
Post-college, Bill has been in the P&C insurance industry as an actuary, product manager, and pricing manager. Bill and his wife Millie are amateur investors with a variety of holdings, but they prefer to buy and hold value investments. In typical "value" style, they live cheap, driving old cars and preferring to save or invest instead of buying fancy "stuff."
Disclaimer: This commentary is solely meant for education purposes and is not intended as investment advice. Please note that the opinions expressed in this commentary are those of the individual author and do not necessarily represent the opinion of MarketThoughts LLC or its management.
TRPITS stands for "Turning a Research Paper Into a Trading System." This is very much in keeping with the theme of my previous MarketThoughts commentary on Using Public Information to Find a Trading Edge. I will briefly summarize the previous installments, then walk through the composition and results from three long/short strategies based on them.
TRPITS, Part I
TRPITS, Part II
TRPITS, Part III
TRPITS, Part IV
The research paper by Andrew W. Lo and Pankaj N. Patel, 130/30: The New Long-Only (PDF), was intended to promulgate an index, or grouping of stocks, that is transparent, liquid, and mechanically implemented, in order that it may be used as a benchmark index for funds of the 130/30 long/short style.
They use the marketing term "alpha factors" to describe the mixture of stock characteristics that they screen for. See Appendix A.1. on page 30 of 44 on the PDF file. They derive 50 indicators from what they label as the "10 Credit Suisse composite alpha factors," and rank stocks from the S&P 500 based on the straight average of their ranking in each of the 10 "factors," each of which is comprised of several indicators.
These 10 "alpha factors" and their 50 indicators boil down to the following seven functional groupings:
(1) "Value" ratios of price to earnings, sales, book value, etc.
_(1.a) applied on an absolute basis across the S&P 500
_(1.b) applied on a relative basis across the stock's industry group
_(1.c) applied on a relative basis across the group's 5-year average
(2) Growth in earnings
_(2.a) on a historical basis
_(2.b) on a future projection basis
_(2.c) changes to analysts' estimates of future projected growth
(3) Growth in sales and cash flows on a historical basis
(4) Assorted fundamental trends in financial statement ratios
(5) Price momentum / relative strength
(6) Short-term mean-reversion technical indicators
(7) Market capitalization (smaller is better)
I have grouped them functionally into seven categories, because I like a functional list with no overlap. In practice, they use ten categories that have significant overlap. They use several of their other supporting products in constructing the ranking, since Credit Suisse already performs evaluation of stocks in the S&P 1500 for their customers, and they use a product from Barra to determine the final weighting of the stocks.
In reviewing the metrics for the various tests performed in the paper, I determined that:
Lesson #1: the restriction on turnover is a constraint on returns.
Lesson #2: there is no significant improvement in returns through the addition of extended long and short exposure in their methodology.
Lesson #3: they like to make things … complicated. Probably more complicated than necessary.
Lesson #4: a more effective extension of the long exposure could theoretically be attained by increasing exposure to the best stocks, rather than increasing the number of stocks with long exposure, and
Lesson #5: a more effective extension of the short exposure could theoretically be attained by increasing the weighting of exposure to each shorted stock, and decreasing the number of stocks shorted.
Retail System Design
The system screens for applicable stocks once every four weeks, making transactions the following day. Test results assume no dividend payments, no slippage, no commissions, and trading at the closing price on the Friday leading into the weekend. Yes, I know those aren't perfect conditions. However, dividend payments would be a positive factor that would counterbalance commissions if the trade sizes were reasonable, and slippage, along with the Friday close/Monday open differential, could work either way (positive or negative) over a long sequence of transactions over many years. Therefore, I consider the conditions reasonable, although imperfect.
ALL tests share the following filters:
Keep :OR([SI Exchange]=”N”,[SI Exchange]=”A”,[SI Exchange]=”M”)
Keep :[SI Market Cap Q1]>100
All LONG tests share the following filters:
Keep :[SI EPS Est Q0-Revisions up]>0
Keep :[SI EPS Est Q0-Revisions down]=0
All SHORT tests share the following filters:
Keep :[SI EPS Est Q0-Revisions up]=0
Keep :[SI EPS Est Q0-Revisions down]>0
The scoring mechanism is calculated this way:
Create [ScoreRank] :[SI % Rank-Cash Flow-Growth 12m]+[SI % Rank-EPS Growth Est]+[SI % Rank-Net margin 12m]+[SI % Rank-Price/Book]+[SI % Rank-Sales-Growth 1yr]+[SI % Rank-Price/Sales]
If Momentum is used in the score, it is represented by the term: + [SI % Rank-Rel Strength 52 week]
LONG tests sort by score in Descending order; SHORT tests sort by score in Ascending order.
In the LONG tests where Momentum is used as both a part of the score, and as a filter, the filter is:
Keep :[SI Price as % of 52 Week High]>90
In the LONG tests where Momentum is represented by two filters, but not in the score, the filters are:
Keep :[SI Price as % of 52 Week High]>90
Keep :[SI % Rank-Rel Strength 52 week]>75
In the SHORT tests where Momentum is used as both a part of the score, and as a filter, the filter is:
Keep :[SI Price]/[SI Price–low 52 week]<1.10
Several SHORT tests were run, but not presented, because their results were not worth presenting.
After reviewing the TRPITS, Part IV results, I have decided to move forward with the Long, Momentum as Two Filters, Hold 30 Stocks "LONG (2F, 30)" and Short, Momentum as a Score, Hold 10 Stocks "SHORT (S, 10)" for my two components.
Notes On Long/Short Composition
Since every mix will be holding 30 stocks long and 10 stocks short, the weights used in test will be composed from those mixes, with rebalancing taking place monthly. Position size is the percent weight divided by the number of stocks. For example, in a 125/25 mix, the average long position would have 125% divided by 30 stocks = 4.17% of equity, and the average short position would have 25% divided by 10 stocks = 2.5% of equity.
To capture the limitations of a retail account, these plans will have a limit of 50% margin, i.e., they will be able to buy and/or short stocks up to 150% of equity. Hence, 125/25 would be the "fully long" equivalent of 130/30.
Further, often a retail investor will be charged margin interest. The assumption with many hedge funds is that the short seller will hold the cash at a risk-free rate of return, making their stream for "equity neutral" equal to risk free rates plus the difference between their long and short returns. In this retail example, the short seller is charged 10% margin interest annually, and gets the risk-free as offset only when the short sale is anchored by cash, and not by stocks. For example, in a 75/75 plan, I am assuming 10% interest debit on the 75% that is short, and a risk-free return on the 25% cash anchor, in addition to the weighted returns of the short and long components.
Presentation of Test Results
CAGR = compounded annual growth rate.
AVG and STDEV = average monthly return and the standard deviation of the returns.
AVG/STD = the average divided by the standard deviation. A risk-adjusted measure.
MAX DD = the maximum equity drawdown. This is evaluated once every four weeks, so it is possible that any system generated greater drawdowns intra-month, but recovered by month-end.
CAGR/DD = the compounded annual growth rate divided by the maximum drawdown. Similar to Seykota's “bliss ratio” but with an annualized return component.
%of Time in DD = how many times, at month-end, was the system in a drawdown, vs. making new highs.
Sharpe, RF=4.5% = the Sharpe ratio with “risk free” being 4.5% annually.
Sortino, MAR=20% = the Sortino ratio with a Minimum Acceptable Return of 20% annually.
The test period is ten years, from August 31, 1997 to September 7, 2007.
The three test portfolios presented are a "balanced" 125/25, an "equity neutral" 75/75, and a 100/33 that assumes the short extension is the better value-added option, considering the 30-stock long portfolio to be already "optimized."
Here are the results presented in terms of equity curve, evaluated monthly.
Here are the results in terms of drawdown curve, evaluated monthly.
Based on the tests that I have run, I would consider the 100/33 portfolio to be my personal choice if I were confined to using one of these systems. That is, of course, a personal preference based on the slightly lower return metrics, combined with a better i.e. lower risk metric for that system.
I do not assume that I am using the best combination of the factors presented in the paper. While I've tried to include the same 10 areas in the same weighting, there may be better i.e. more optimal solutions.
The factors chosen, and parameters used for the filters, may not be optimal.
A holding period of four weeks may not be optimal. Perhaps shorter, or even longer, periods might work better.
Some of the long ideas presented in TRPITS, Part IV may actually be better ideas if either (1) used in conjunction with a market timing system, or (2) used as a filter for a chart-based discretionary trader to make decisions based on.
I believe that the final system above is a workable trading system suitable for an intelligent retail trader. It is comparable, in backtested performance, to many of the systems that I tested in the development of the systems I am tracking at The Rempel Report. That, however, is not the point.
This is an exercise based on the Lo/Patel paper, and isn't necessarily the best system, it isn't necessarily the best use of the material in the paper, and it isn't necessarily the optimal long/short result.
My objective here has been to show how a retail trader can take knowledge that is in the public domain and construct reasonable, realistic trading systems that are both mechanical in nature, and outperform the broad market by quite a bit. The knowledge and information necessary to trade successfully is available, and in most cases, is free.
I do recognize that trading one's own account involves more than just knowledge of the systems that work, or the underlying characteristics of stocks that, statistically speaking at least, contribute to outperformance. The essence of trading is finding an edge, applying it repeatedly, and using good risk control. Work such as that performed above provides information on the edge, a diversified approach that mitigates individual company risk and seems to produce lower drawdowns, and a structure that applies the process repeatedly. Left out of the equation so far has been a focus on the emotional and mental aspects of trading, which can make use fail to consistently apply that edge, or lose our concentration on control of risk.
More on this later.