For investors, experience is the best teacher -- even for a computer-driven stock-selection strategy. That's the basic approach of Standard & Poor's Neural Fair Value 25 portfolio, which employs the investment research outfit's proprietary quantitative stock ranking system.
The Neural Fair Value (NFV) concept, which was created by Andre Archambault, S&P's director of quantitative strategies, starts with S&P's Fair Value stock valuation system, which uses earnings estimates and other metrics to determine whether stocks are trading above or below their fair value. The "neural" part comes into play when Archambault's model, updated weekly, combs the 3,000 stocks in that group for the 25 names it thinks have superior price appreciation potential. Some notable current members of its Neural Fair Value 25 portfolio: Anadarko Petroleum (APC), Black & Decker (BDK), Hewlett-Packard (HPQ), Home Depot (HD), Lockheed Martin (LMT), Northrop Grumman (NOC), and RF Micro Devices (RFMD).
PAST AS PROLOGUE. The NFV approach, Archambault explains, is based on "Neural Network" theory, an artificial intelligence concept that seeks to replicate the human brain's ability to learn from mistakes. Just like flesh-and-blood stock pickers, a Neural Network has to undergo a "training period" in which key inputs are adjusted to reduce the model's prediction errors. In S&P's neural model, Archambault inputs the factors that produced better market performance in stocks during the most recent six-month period to project those that will prosper in the coming six months.
The NFV 25 portfolio has handily beaten its benchmark, the S&P 500 index. Year-to-date through Apr. 28, it was up 5.4%, vs. a 5.0% rise in the S&P 500. Since its Oct. 5, 2004, inception, the NFV 25 has risen 30.7%, vs. 15.5 % for the index. S&P notes that these are model results only and do not represent the performance of actual trading of assets, and that the portfolio would be subject to considerable risk given the volatility and turnover inherent in the model.
BusinessWeek Online's Will Andrews spoke with Archambault about the inspiration for the NFV 25 portfolio and how his stock-selection system works. Edited excerpts from their conversation follow.
What was the inspiration for the Neural Fair Value concept?
Before, we just created portfolios with our Fair Value rankings alone, including an earnings surprise model. We found in the late 1990s that there was an awful lot of reliance on company guidance for the earnings estimates [used in the Fair Value model], and as a result, we ended up with a lot of companies with high earnings surprise ranks. The companies were low-balling earnings so they could beat them all the time.
There was also a mandate to create a new portfolio that was restricted to a limited number of stocks.
At the same time, I figured I could see what else we could do to make the portfolios more sensitive to current market conditions -- and less sensitive to what the companies were doing about managing their earnings and guidance. That's when I started working with the neural networks.
How do you determine the starting point, fair value?
Fair value is basically driven by consensus earnings forecasts for one-year, two-year, and five-year earnings growth. The rationale for that is basically that expected earnings are what drive stock prices. There are a lot of internal calculations relating this to the company's history over the past 10 years - the minimum necessary is 10 quarters. We compare those with the similar data points of the peer group of these companies and also all the companies in the S&P 500 index. And then we do multiple regressions and it compares all the fair value numbers that are derived by these different comparisons and we end up with one final number, the fair value.
We compare the fair value with the current price, and we get an expected return. We rank all these expect returns from the very highest to the lowest and we place them in quintiles. Those give us our fair value ranks, ranging from 5, highly undervalued and most attractive, to 1, highly overvalued and least attractive.
The Fair Value concept is familiar to many investors, but the part that makes this unique is the neural overlay. How does the artificial intelligence concept come into play?
Neural nets are kind of like "black boxes," and they're being used in all kinds of industries. They're being used to optimize whatever people are doing -- if it's in insurance, to minimize risk, or in data processing, to maximize processing power and so forth.
What I'm trying to do every week is go back about six months and look at what happened to each individual stock in the Fair Value universe during that period. I compare the price of each stock to its price at that time and I calculate how it did during those six months. Then I group them into quintiles and I end up with a quintile number, 5 through 1.
So that's basically the goal -- the target -- that I would like to emulate. I take the target number and I feed it to the "black box" and it uses all the fair value inputs and outputs that were generated six months ago, and from that it tries to emulate the numbers that resulted in the following six months.
Then I bring that model to today's date, feed to that model the current fair value numbers, and tell the model to give me a new target number for the next six months. It will not be perfect, because market conditions can change, but it assumes that what worked in the past six months will work again in the next six months. And it gives me the number [1-5] as a classification to help me better time the buys and sells of the stocks in the portfolio.
So like people, the model "learns" through the accrual of experience.
Yes, every week it goes back six months and if it worked in that time, we'll make a little leap of faith that it'll work, in general, for the next six months.
I [recently] started thinking I should use another model that will not only look at what happened six months ago but will look at what happened to each stock every single month and bring it up to date. This is something we are about to integrate and combine with the current fair value classification so that we have it both ways, and they are even more sensitive to what is occurring between six months ago and six months from now.
Where can individual investors find these portfolios?
The Neural Fair Value 25 can be found on S&P's Advisor Insight or Outlook.