DATA BEATS THE S&P 500 – Decision Rule I

Deception And Truth Analysis (D.A.T.A.), Inc. has spent almost a decade working with deception scientists[1]and studying multiple methods of deception and truth detection. Based on this knowledge we developed Deception And Truth Analysis, a proprietary algorithm based on scientific research and making use of textual data to rapidly identify deceptiveness and truthfulness in business communications.

A practical demonstration of its power is our recent work where use of D.A.T.A.’s algorithm was able to best the performance of the Standard & Poor’s 500 Index. Using the 500 components of the S&P 500 we created a simple decision rule: at the start of each new year simply do not buy the 50 firms (10% of firms) whose level of deception was assessed as highest in the previous year. More specifically the methodology is, as follows:

1. Starting with 2008 and continuing through 2021 each of the 500 S&P 500 components’ 10(k)s were assessed by D.A.T.A. to create a D.A.T.A. Score for each firm. DATA Scores range between -100% and +100% with any negative score indicating a level of deceptiveness and any positive score indicating a level of truthfulness, in the aggregate.

2. Then, on the first trading day of the subsequent year we purchase the 450 S&P 500 components with the highest D.A.T.A. Score (i.e. those assessed as most truthful), and do not purchase the 50 S&P 500 components with the lowest D.A.T.A. Score (i.e. those assessed as most deceptive). Each of the 450 purchased companies is equal weighted in a portfolio and held for the entire year. The average lag from assessment by a D.A.T.A. Score to purchase is 9.08 months.

3. During 2009 each of the 500 S&P 500 companies issue annual 10(k) reports again which are assessed by D.A.T.A. Scores.

4. On 31 December 2009 the portfolio of most the truthful 450 equal-weighted S&P 500 components from 2008 is sold.

5. On the first trading day of 2010, we purchase the 450 S&P 500 components with the highest 2009 D.A.T.A. Scores (i.e. the most truthful), and do not purchase the 50 S&P 500 components with the lowest D.A.T.A. Scores (i.e. the most deceptive). Each of the 450 purchased companies is equal weighted in a portfolio. This is the same as in step 2, above.

6. This technique of D.A.T.A. Scores being generated for each S&P 500 components’ 10(k) documents, then waiting until the start of the next year to buy the prior year’s 450 firms assessed as most truthful/having the highest D.A.T.A. Scores, is repeated through until 31 December 2021.

Note: A separate analysis was done where instead of selling the portfolio of 450 names on 31 December each year, they are held for 5 years. But the methodology is identical otherwise.

Here are the results of our simple decision rule just described:

1. Where the strategy is to not buy the 50 firms with the lowest D.A.T.A. Scores/those assessed as most deceptive during the year:

• • Annual average outperformance of 56 bps for a compounded total return advantage of 6,243 bps (2008-2021), with a standard deviation of 0.46%. This comparison is our equal weighted 450 holdings vs. equal weight S&P 500. Note: results against the actual S&P 500 are quoted below; and they are much higher.
  • Outperformance in 11 of 13 years (2008-2021).
  • Worst performance is -19 bps in 2020.
  • Best performance is +133 bps in 2019.

2. Where the strategy is to not buy the 50 firms with the worst D.A.T.A. Scores at the start of each year and hold for 5 years:

• • Average rolling 5-year outperformance of 566 bps, with a standard deviation of 0.97%.
  • Outperformance in 9 of the 9 five-year rolling periods.
  • Worst performance is +367 bps for the five-year period 2013-2017.
  • Best performance is +681 bps for the five-year period 2010-2014.

3. If a direct comparison is made of our strategy against the actual S&P 500 returns (i.e. not equal-weighted S&P 500), the results are, as follows:

• • Average annual outperformance of 895 bps for a compound total return advantage of 71,060 bps (2008-2021), with a standard deviation of 26.83%.
  • Outperformance in 7 of 13 years (2008-2021).
  • Worst performance is -2,789 bps in 2017.
  • Best performance is +7,447bps in 2008.
  • <Aside: equal-weight is a badass strategy.>

4. When dividing the S&P 500 components into performance groupings each year (2008-2021) by deciles, quartiles, and quintiles, we find monotonic returns for the first 300 companies excluded, as follows:

DECILES

• • 33.58% (1-50 most deceptive companies excluded)
  • 84.38% (51-100 most deceptive companies excluded)
  • 138.75% (101-150 most deceptive companies excluded)
  • 171.60% (151-200 most deceptive companies excluded)
  • 206.97% (201-250 most deceptive companies excluded)
  • 258.16% (251-300 most deceptive companies excluded
  • Thereafter the return advantage steadily declines as the most truthful companies are excluded from purchase.

QUINTILES

• • 58.98% (1-100 most deceptive companies excluded)
  • 55.17% (101-200 most deceptive companies excluded)
  • 232.57% (201-300 most deceptive companies excluded)
  • Thereafter the return advantage steadily declines as the most truthful companies are excluded from purchase.

QUARTILES…monotonic returns for the first 375 companies excluded:

• • 72.78% (1-125 most deceptive companies excluded)
  • 81.34% (126-250 most deceptive companies excluded)
  • 241.01% (251-375 most deceptive companies excluded)
  • Thereafter the return advantage steadily declines as the most truthful companies are excluded from purchase.

5. The results reported in #1 in our summary above are first pass results and are not optimized in any way. An example of optimized results would be to relax the number of holdings not purchased from an arbitrary 10% and to instead exclude the optimal number of companies whose 10(k)s had the worst D.A.T.A. Scores to maximize the return advantage of the decision rule. If we did that the results would be to exclude the 353 firms whose DATA Scores were the lowest.

• • Annual average outperformance of 250 bps (2008-2021) vs. 56 bps for the simple decision rule of excluding the 10% of companies whose DATA Scores are the worst and whose results are shared in #1, above. Excluding the 353 firms whose D.A.T.A. Scores are lowest results in a compounded total return advantage of 31,342 bps (2008-2021) vs. a non-optimized 6,243, with a standard deviation of 3.58%.
  • Outperformance occurs in 10 of 13 years (2008-2021).
  • Worst performance is -311 bps in 2018.
  • Best performance is +932 bps in 2019.

Outperformance of such magnitude consistently occurring for multiple time-periods, with an average time-to-purchase lag of 9.08 months, in some of the most liquid and covered stocks in the world is noteworthy. Additionally, the DATA Scores utilized were for the entire 10(k) and not just the more extemporaneously authored MD&A section.

Note:

• No other analyses were applied, just a run of each S&P 500 component’s MD&As through D.A.T.A.’s algorithm.
• The D.A.T.A. algorithm was developed out-of-sample and based on how people pan culturally deceive using language, and not developed in-sample based on how businesses deceive through language.
• To avoid overfit, the main reported results are not optimized.
• There is a considerable advantage to equal-weighting, but the scores we report are “apples to apples,” or equal-weight S&P 500 compared to equal-weight S&P 500, less the 50 most deceptive firms.

We would of course love to hear from you so that we may demonstrate Deception And Truth Analysis (D.A.T.A.) technology for you and demonstrate the many practical uses beyond alpha-generation for you and your research team.

[1]See Jason Apollo Voss CFA’s Lie Detection Guide: Theory and Practice for Investment Professionals, for example.