Previous research suggests that there exist strategies that, when implemented over a long period of time, might provide higher returns than overall market performance (see, e.g. [1]). One of these strategies, namely the “Max-Median Rule”, was investigated by Thompson and Baggett (see [2]), and served as a general motivation for this research. By selecting a handful of stocks, according to some robust criterion (e.g. the median) and rebalancing consistently without straying away from the strategy, virtually any investor could easily manage his or her portfolio quite reasonably. Over the long-haul, this strategy would provide decent returns when compared to a benchmark index (e.g. the S&P 500 Index). It is worthwhile noting that in strategies such as these, time is a major consideration (and one which investors can control, e.g. when investing retirement funds such as a 401K401K), and that these methods do not constitute day-trading strategies, and should be adhered-to consistently over a given period.

Several salient points of this motivating investment strategy are:

These points clearly serve us as a motivation for further investigation and potential improvements. In particular, through recognizing that the existing strategy, albeit well-performing, is inherently a “one-at-a-time” strategy and therefore does not capture any correlation-related dynamics through its selection criteria.

Lastly, we were also motivated to investigate (at least initially) equally-weighted portfolios. An interesting finding (see, e.g. Wojciechowski, Baggett, and Thompson [3]), is that “for the 33 years from 1970 through 2002, not simply a flukish few, but a staggering 65 percent of the portfolios selected randomly from the 1,000 largest market cap stocks lie above the Capital Market Line (CML).” Also, it has been shown (see [2]) that any individual who invested equally in the S&P 500 Constituents (time period of 1970 through 2006) would have made, on a yearly average, 13.7%, as opposed to 8.9% with a competing market-cap weighted strategy. Both of these empirical realities make, at least preliminarily, a case against considering long-term market-cap weighted strategies.

We now consider a strategy which allows us to implicitly capture the joint performance of securities as part of our selection criteria. Our goal is to pick, from the universe of investible stocks, a meaningful handful on which to equally allocate a given investment quantity on a yearly basis. As a first step, we consider the S&P 500 constituents to be our universe of stocks from which can select a critical few (a number which we have set initially, and somewhat arbitrarily, to 20. This quantity seemed both appealing and reasonable in terms of being financially manageable and computationally feasible). It is also worthwhile noting, that we regard limiting an investor to select from the S&P 500 (or any other well-known index) as both a reasonable and soundly restricted starting point. Furthermore, we also know that stocks listed in the S&P 500 are representative of various market sectors (inherently diversified) as well as of various reasonable company sizes (in terms of market capitalization). Additionally, other filtering criteria inherent in a reasonable-size index (in terms of the number of constituents), seem to provide a good baseline both as a benchmark (to outperform) and as a sensible constraint to the universe of all potentially-considered stocks.

Our first step is to select a subset of stocks from a given index in which to allocate a given investment at any given point in time. Here, and in general, we can start by considering a subset of n stocks from a given index I with K constituents. Our evaluations considered I=S&P 500 (for which K=500K=500) and assembling baskets of n=20n=20 each time. Based on this setup, there is a total of C20500≈2.667×1035C20500≈2.667×1035 unique baskets of randomly selected securities that we could potentially consider. Clearly, if we require evaluating some optimal objective function over all possible combinations, this becomes computationally infeasible.

Instead, we proceed by selecting stocks according to some plausible robust criterion that can be applied to any randomly assembled basket (the most appealing, both in terms of interpretation and prior results, being the median of the portfolio daily-returns). We also note, and quite emphatically so, that to both evaluate a meaningful amount of portfolios as well as to assess procedure repeatability, we clearly need to parallelize this effort.

Consider the following algorithm:

We pick a subset of n=20n=20 yearly “investible” stocks according to the aforedescribed criterion at the end of any given year (using the most recent one-year data). We then allocate our investment quantity in this portfolio, on the first trading day of the subsequent year, holding it for one year and concurrently collecting data during this year to repeat this procedure at the end of the year. We essentially keep on repeating this procedure over the period for which we want to evaluate the strategy.

It is also interesting to note that under the previous motivating rule (i.e. the Max-Median Rule), we would always get an exact answer regarding which stocks had the single highest max-medians, in a finite, rather short, amount of time (an essentially NPNP-complete problem). This implied determinism, in the sense that any subsequent runs would produce the same results, amounts to a variance of zero. However, it is rather evident that our modified algorithm is inherently stochastic as we cannot evaluate all possibly imaginable combinations of portfolios. As a direct consequence, and by randomly selecting a reasonable number of portfolios for evaluation we expect to observe some natural variation, in the sense of the procedure's repeatability (each run will be essentially unique). It is possible (and rather interesting) to exploit this natural variation to assess the overall repeatability of this modified procedure.

Data were obtained from the University of Pennsylvania Wharton/WRDS Repository [4]. The following data were utilized for our evaluations:

Data were also obtained from Yahoo! Finance:

For our evaluations we note that our yearly returns with dividends were calculated from the first trading day to the last trading day per year and that dividends were included. Also the size of the data files analyzed was approximately 900MB.

It is worthwhile mentioning some general details regarding the overall parallelized implementation of this procedure. It was successfully implemented using the software R, widely and freely available from the Comprehensive R Archive Network (CRAN). Several packages available for R, make a parallelized implementation of the algorithm very straightforward. In particular, we made use of snow (see, e.g. [5] and [6]), and snowfall (see [7]), both running over open-MPI. Some of the reasons for choosing this implementation were:

In essence, this approach was very appealing in terms of performance, development time, and cost (essentially free). Although faster and more efficient implementations are possible (e.g. C/C++ and Fortran with open-MPI, the aforementioned implementation was sufficient for our purposes).

The code utilized was initially developed for sequential execution (in SAS) and then converted to R with similar performance. It was subsequently converted from sequential to parallel to exploit the benefits of a parallel R-implementation. The standard steps for this conversion process are pretty standard, essentially:

Several 64 processor jobs were submitted to Rice's Cray XD1 Research Cluster (ADA) - occupying 16 nodes, each with a total of 4 processors. The jobs would take less than 20 hours to complete, for about 64 simulated tracks of 40 years each.

Several simulations were run with observed, above average performance (number of portfolios inspected per simulation was in the range of J=25,000J=25,000 to J=50,000J=50,000). Figure 1 shows a simulation of J=50,000J=50,000 portfolios per year over a period of 43 years. Several interesting features can be noted through this figure. We can appreciate that any investor that made up a portfolio with an initial investment of $100,000 in 1970 and selected the same stocks chosen by our algorithm would have allowed his or her portfolio to compound to a total of $3.7M (by the end of 2008), which performed better than both an equal-investment strategy in the S&P 500 Index (about $2.7M) or a market-cap weighted investment strategy in the S&P 500 Index (slightly below $1M). Of course, we can imagine that the computational power that was used was not available in the early seventies, but moving forward it will be and to an extent this is what matters. Also it is clearly seen that the Coordinated Max-Median Rule is inherently a more volatile rule (as compared to the S&P 500). Next, Figure 2 describes 1 of 3 pilot runs that were evaluated with various measures suggesting the superiority of the max-median (as opposed to, say for instance the mean, as well as the benefit of using the “max” rather than the “min”). This seems plausible, at least heuristically, as the median is most robust (the mean is least robust), and in some sense the previous years “best” performing companies are more likely to perform better next year than the previous years “worst” performing companies (in terms of returns).

Figure 1: Coordinated Max-Median Rule (Single Run) 50,000 Portfolios Evaluated per Year

Figure 2: Coordinated Max-Median Rule (Single Run) 25,000 Portfolios Evaluated per Year with additional evaluations

Recent evaluations have been mostly focused on evaluating the following:

Several experiments were set up to determine if it would be worthwhile to inspect more randomly sought portfolios on a yearly basis as part of the overall procedure. A job simulating a total of 104 tracks (each consisting of J=25,000J=25,000 portfolios per year over a 43 year period, 1965 though 2008) was submitted to ADA and took approximately three days to complete. Several important observations can be made from the outcomes of these simulations (shown in Figures 3 and 4, below). We note that, here, we can exploit the independence regarding the portfolios evaluated to get 52 tracks of J=50,000J=50,000 portfolios each by combining pairs of J=25,000J=25,000 tracks and selecting the maximum of the pair (simply the maximum of a longer execution). Essentially this gives us information regarding what would have happened (in terms of the performance of the strategy should we have run it for twice as long). Analogously, tracks for J=100,000J=100,000 and J=200,000J=200,000 portfolios were constructed. Finally, some overall discussion of the results is given after the figures.

Figure 3: Procedure Repeatability and Number of Portfolios Sampled Simulation (Left: J=25,000J=25,000 tracks, Right: J=50,000J=50,000 tracks)

Figure 4: Procedure Repeatability and Number of Portfolios Sampled Simulation (Left: J=100,000J=100,000 tracks, Right: J=200,000J=200,000 tracks)

The total portfolio value was evaluated at the end of both years 2006 and 2008 and contrasted to both market performance (blue track) and the performance of a single track of a whopping J=2,600,000J=2,600,000 portfolios considered yearly (green track). As expected the variability of the procedure compounds as a function of time, and by chance we might under-perform the market. However, more often than not the procedure out-performed the market and by a quite reasonable amount. The proportion of portfolio-tracks simulated that were over an equally-weighted alternative at the end of 2006 was over 80% (for the cases where J=25,000J=25,000 and J=50,000J=50,000) and over 90% (for the cases where we assessed more randomly sought portfolios, i.e. J=100,000J=100,000, and J=200,000J=200,000). Also, there is weak evidence suggesting that, although running as many as J=2,600,000J=2,600,000 portfolios might at times outperform the market, this approach is generally not consistently higher on-average than considering tracks consisting of less yearly-evaluated portfolios.

Another rather interesting observation is made through the scatter-grams produced (see Figure 5, below) assessing the correlation between current year portfolio median and (same portfolio) next year performance contrasted to the performance of the S&P 500 index. The number of portfolios evaluated for this purpose was 2,000,0002,000,000 and the those that are highlighted as producing the maximum of the medians represent (<0.1%<0.1%, i.e. <2,000<2,000). The main purpose of this effort was to assess any associations between the current year medians as a forward-looking measure of portfolio performance (as we intend to pick the maximum and by chance we can pick portfolios of performance similar to those in the top 0.999 percentile). As expected the associations are weak, though not extremely weak (correlations are 0.2090.209 for the first case and -0.182-0.182 for the second), however can be noticed and depend highly of the year evaluated.

More often than not, we observed a positive correlation for the years inspected (the strongest correlations are those shown in the figures below). It turns out that for certain years (those with a negative correlation), we ought to utilize the “min-median” as a selection criterion. However, this cannot be known ex-ante, and the best we can do is utilize a measure that more often than not, produces above-average results. Here again, we can appreciate how these conflicting effects would average-out with time in a favorable direction, reiterating the fact that a strategy such as this one, if considered, should be evaluated over the long-haul.

Figure 5: Next Year Portfolio Performances vs. Current Year Portfolio Medians (Left: 1998 vs. 1997, Right: 2001 vs. 2000)

Lastly, several evaluations were performed comparing the various max-medians of the portfolios simulated as a function of the number of portfolios run (i.e. J) and compared to the single-stock max-median (See Figure 6 below), which could, at least heuristically, serve as an upper bound. This resulted (empirically) to be somewhat unstable as there is no guarantee that any thresholds set in terms of percentage to the bound could be attained in any reasonable computing time, mainly due to the fact the after a reasonable amount of simulations (namely J=500,000J=500,000 and up to J=2,000,000J=2,000,000) the percentages of this single-stock max-median attained depended considerably on the year inspected, making a generalization impossible. The most recent evaluations were performed with stopping after 5 ticks past J=10,000J=10,000 simulations, which seems stable, however based on aforementioned results it seems to not provide any incremental benefit when contrasted to, for instance a hard-coded constant J stopping rule.

Figure 6: Max-Median Searches as a function of J (Left: 1984, Right: 1991)

Several items are open at this point that might be worthwhile investigating in future research. Amongst them are the following (to mention a few):