Altius Minerals (Toronto: ALS) is one of our favorite junior resource companies; the miner is pursui...
Provenance and Promise: Restoring Robust Life to Powerful Indicators
11/15/2016 5:07 pm EST
I’ve been developing stock market timing models since the early 1980s, almost thirty years of effort. In that time I’ve drawn some conclusions about the logic and structure of systematic investment strategies, findings I’d like share with you even if your personal perspective on this matter differs.
First, long-term models that hold positions for weeks or months tend to outperform shorter-term strategies that trade several times a week or even more frequently. This means that weekly models are generally more reliable than daily or intraday mechanical timing methods. (Monthly models are too deliberate for my taste with the exception of Meb Faber’s inspired asset allocation portfolio.)
Second, there are many indicators with forecasting significance for the stock market. In past studies we have featured monetary variables like interest rates and credit spreads as well as breadth, sentiment and other leading indicators. The bad news, at least from my experience, is that most predictive inputs are likely to fail at some point in the future. Even if an indicator’s track record has been flawless for years or decades, it is apt to stumble. As a result, you’ll either be dangerously over-exposed on the long side when the market tanks or you’ll be stranded on the sidelines when the market surges higher.
A good example of a potentially misleading predictive variable is interest rates. When I started Formula Research in 1991, the professional consensus on the treatment of monetary data was simple. If treasury yields went down, that was good for stocks. If treasury yields went up, that was bad. But we learned a very different lesson from Japan in the 1990s and the U.S. in 1998 and in the recent financial crisis. In periods of economic contraction, falling interest rates can presage or coincide with sharply lower share prices. Once rates finally hit bottom, rising yields can actually be welcome development for both the economy and the stock market.
My point is that fundamental and other predictive indicators can operate reliably for a long time only to falter at some unpredictable point in the future. The losses may be few and far between but when they occur they can be devastating. So what can we do to solve this problem? I recommend three partial solutions (there is no one conclusive remedy). First, try to track as many forecasting methods as you can manage. The idea is to diversify your mix of screening tools so that if any one indicator falls short--as is nearly inevitable--other complementary but distinct methods will help compensate.
Second, supplement your roster of forecasting indicators with some purely trend-following signals. These are price-based strategies that don’t try to anticipate financial trends but merely seek to track and exploit them. Percent swing reversals and moving average crossovers are examples of purely trend-following entry and exit signals.
We’ve published a number of trend-following models for stocks over the years. Four of them are particularly relevant to the present discussion. As it happens, all four were published in one year, 1994. The Fabian model (Jan-94) has returned 11.07% compounded annually since 1970 compared to 9.91% for the S&P 500. An initial $10,000 investment in 1970 would have grown to $684,600 compared to $448,600 for S&P total return.
The Super 248 model (Feb-94) returned 12.09% over the same period. An initial $10,000 stake in 1970 grew to $991,900 today. The November 1994 issue actually featured two trend-following models. The Allen model gained 11.09% compounded since 1970. A $10,000 investment would by now have grown to $691,300. And the Traub Trend model returned 11.46% over between 1970 and early 2010. A $10,000 stake would now be worth $789,700.
The dollar gains posted by the four models just cited outperformed the S&P 500 by 53%, 121%, 54% and 76% respectively. With these kinds of returns, why not just rely on trend-following strategies and ignore predictive indicators altogether? The reason is that purely price-based models are not very accurate. The percentage of winning trades typically ranges from 40% to 50%. While individual losses are minimal, collectively they can exacerbate risk. The upshot is uncomfortable levels of cumulative drawdown.
One recommended fix will become the defining theme of this report and is treated in ample depth below. Another solution is the same tactic cited above, diversification. If you blend a variety of trend-following indicators into a composite, you’ll get a whole greater than the sum of its parts. The upshot is higher returns and lower risk than any single trend-following tool.
As an example, some time ago I prepared a custom composite for an institutional money manager in Hong Kong. While the model factored in key predictive indicators, the four trend-following strategies were vital to the logic. That model has returned almost 16% a year since 1970 with drawdown held to 13%. I developed the model about ten years ago, just ahead of one of the most challenging periods in U.S. financial history. Thanks to those four trend-following indicators, the custom composite weathered the storm ably.
So let’s sum up. We find predictive models appealing because they tap into the causitive forces that shape and determine price behavior. The problem is that even the best forecasting tools occasionally trip up. When they do the damage can be devastating. On the other hand, trend-following models tend to trigger frequent false signals. In isolation the losses are modest, but over time they can promote undesirably high risk.
The SVI: How to Fix a Broken Predictive Model
In late 1992 I devoted three reports to one of the most innovative stock market indicators I had ever seen. The SVI stands for Stock Valuation Indicator. The model was developed by Peter Martin and Byron McCann, analysts and money managers who introduced the formula in their classic text, Investor’s Guide to Fidelity Funds (Wiley, 1989).
The SVI is a classic predictive indicator relying exclusively on fundamentals. In the December 1992 report, your correspondent wrote excitedly: “I have always sought out indicators that reach beyond the domain of price to capture valuation fixed to some external standard.” The SVI seemed ideal for this mission.
The idea behind the SVI is appealingly straightforward. You take two measures of stock market valuation--the dividend yield and the earnings yield--and compare them to the 90-day T-bill rate. A high ratio means stocks are undervalued relative to T-bills--bullish. A low ratio means stocks are overvalued compared to T-bills--bearish.
I won’t spend any time explaining how Martin and McCann derived the SVI. For background consult their fine book or our own Dec-92 issue. Let’s get right down to the simple mechanics of the SVI. Take the S&P 500 dividend yield and multiply it by two. Now average that number with the S&P earnings yield. Finally, divide that result by the 90-day T-bill yield. The formula is: ([2 * DivYld + EarnYld]/2)/T-billYld. Readings above 1.00 are bullish; readings below 1.00 are bearish.
Until the mid-1990s this indicator had a remarkable track record. Between 1970 and 1995 the SVI returned 16.39% compounded annually compared to 11.76% for the S&P 500. A $10,000 investment in 1970 would have grown to $518,600 by late 1995. That’s close to three times the S&P’s dollar gains of $180,400 over the same period. Thirty out of 36 trades were winners, an 83% batting average, not bad for a model that doesn’t directly track price. The chart below shows the primary SVI buy and sell signals from 1970 through 1995. [NF1]
That exemplary track record would soon be shattered. In March 1996 the SVI turned prematurely bearish. The sell signal came just ahead of a four-year, 155% stock market rally. Once a predictive indicator like the SVI generates a signal, it can remain hopelessly locked on to the position for years. In this case the SVI did not issue an offsetting buy signal until August 2001.
You might call this untimely sell signal a false negative. The SVI later succumbed to the opposite error. It gave a buy signal in July 2007 just before the market began its historic plunge. You could call this signal a false positive. The SVI suffered the same 55% drawdown as the SAP 500. (The buy signal remains active today.) The upshot is readily apparent in contrasting performance over two intervals. The SVI made 16.4% a year from 1970 through 1995. By contrast, from 1996 to date the SVI returned 4.73% a year. (The SAP made 11.8% and 6.6% a year respectively.)
The SVI is a great example of a potentially useful fundamental indicator gone awry. The false signals are rare but punishing. Consider. From 1970 to date, the original SVI returned a very respectable 12.12% a year compared to 9.91% for the SAP. In the process the SVI recorded 34 winning trades out of 43 entries, an impressive 79% batting average through forty years of testing. If there were some way to mitigate the destructive impact of the occasional but deadly false signal, it might be possible to breathe new life into the SVI.
At the close of the last report I outlined one possible way to do this. The working template was my version of the Up Sheen earnings yield model. As you may recall, that model requires both a fundamental factor (earnings yield) and a trend-following signal (SAP moving average) to trigger a long entry. A sell signal by either indicator mandates a market exit.
When I adapted the Up Sheen logic to the SVI, results were very promising. Before getting into details let me first describe the rules of the revised SVI. Calculate the SVI formula as described above. When the SVI is greater than its 50-week simple moving average and the SAP 500 is greater than its 50-week moving average, enter long. When either indicator drops below its 50-week moving average, exit to commercial paper.
Since 1970 the new SVI returned 10.64% a year compared to 12.12% for the old SVI and 9.91% for the benchmark SAP 500. The return fell short compared to the old SVI (12.12%).Â But the new SVI also greatly reduced risk. Drawdown was 8.7% compared to 55% for the old SVI, an 85% decline. The 8.7% drawdown is the lowest for any stock market model Formula Research has ever published. Forty-one of 54 trades were profitable, a 76% winning percentage, very high for a trend-sensitive strategy.
Two features that are built into the very structure of the new SVI (and of course its Up Sheen prototype). First, in case the predictive component (the SVI) flashes a false buy signal, you cannot actually go long unless the SAP shows clear evidence of price strength. Second, you will exit the market if either the SVI or the SAP turns negative. This means you will occasionally miss out on a stock market rally that is not confirmed by the SVI.
The opportunity cost is worth it. The gains you forgo are more than offset by the reduction in risk. In the past I have described hybrid logic like the new SVI as “adaptive trend-following.” This name still seems fitting. The chart below shows comparative equity curves for the new SVI ratio and the SAP 500.
How to Fix a Second Broken Predictive Model
In its heyday the T-bill/Dividend ratio was far better known than the SVI. We featured this indicator in our second issue (Oct-91). I was so confident in this method that I highlighted the ratio one year later, in a report that would serve as the free sample newsletter we send out to prospects. [NF2] [NF3] My enthusiasm for the indicator was not dimmed when noted market analyst Ned Davis told me at the time it was his favorite stock market indicator (out of the thousands he tracks).
I’ll explain the logic of the T-bill/Dividend ratio shortly. But when I prepared that Oct-91 study, its track record seemed truly exceptional. Between 1946 and late 1991, the model gained 14.16% compounded annually compared to 11.31% for the SAP 500. An initial $10,000 investment in 1946 grew to $4,344,100 by late 1991 compared to $1,349,700 for the SAP 500. A striking 23 of 26 trades were profitable (88%). There were just three losing trades and they were minor--2.2%, 3.1%, and 0.7%.
Then, as with the SVI, catastrophe struck. The T-bill/Dividend ratio issued a sell signal in November 1994. No buy signal would be emerge for another seven years. By then stocks had surged 134%. As with the SVI, a false negative was aggravated by a false positive. The model gave an ill-timed buy signal in late 2007 right before the worst market decline since the Depression. From late 1991 to date, the T-bill/Dividend ratio returned just 5.26% a year versus 8.49% for the S&P 500, a stark contrast with its previous edge over S&P buy-and-hold.
As you might surmise, the T-bill/Dividend ratio is somewhat similar to the SVI. There are three key differences. First, the ratio tracks only T-bill and dividend yields; the earnings yield is ignored. Second, the T-bill yield is not double-weighted as with the SVI. Third, in this application the numerator and denominator are reversed as compared to the SVI. To calculate the ratio, you divide the 90-day T-bill yield by the S&P dividend yield. As a result of this flip, low readings are bullish and high readings are bearish, the opposite of the SVI.
I first encountered the T-bill/Dividend ratio in a Wall Street Journal article featuring the work of investment manager Richard Eakle. [NF4] Following Eakle, for testing purposes I set the buy and sell parameters at a reading of 1.60. You buy when the ratio drops below 1.60; you sell when it rises above 1.60. [NF5]
So what can we do to redeem the T-bill/dividend ratio? You guessed it...the same thing we did with the SVI. We use adaptive trend-following logic. The rules are the same as the new SVI but with an adjustment to reflect the flip in the numerator and demonitor. In this case you go long when the T-bill/Dividend ratio is below its 50-week simple moving average and the S&P 500 is above its 50-week smoothing. You sell stocks when the T-bill/Dividend ratio is above its 50-week average or the S&P is below its 50-week average.
Since 1970 the revised T-Bill/Dividend model returned 11.42% compared to 9.91% for the S&P 500. A $10,000 investment in 1970 would have grown to $778,100 compared to $448,600 for the S&P. Drawdown fell to 8.9%, the second lowest such reading ever posted in these pages. Forty-two of 56 trades were profitable, a comfortable 75% batting average. The chart below shows comparative equity curves for the new T-Bill/Dividend ratio and the S&P 500.
In this and the previous study we examined three stock market timing models with the same basic structure. You enter long when some valuation measure is positive (earnings yield, SVI, T-bill/dividend ratio) and the S&P 500 is above its moving average. You exit when the valuation input turns negative or the S&P crosses below its moving average. All three models use a 50-week simple moving average, a time span I chose arbitrarily. Yes, you could optimize the smoothing interval for each application individually or collectively for all three. But why bother? Performance is solid without the hassle and the hazard of excessive curve-fitting.
The Pu Shen earnings yield model, the new SVI and the new T-Bill/Dividend ratio all returns far superior to S&P buy-and-hold. Just as important, each strategy exhibits exceptionally low risk. All three models posted unprecedented single-digit drawdowns, making these the three most most risk-averse stock market strategies we have ever published.
The Pu Shen model tracks earnings yield, the T-bill/Dividend ratio tracks dividend yield, and the SVI tracks both. Obviously, with this high degree of association there is going to be some convergence in performance. I haven’t done the calculations but the three equity streams are bound to be highly correlated.
The main reason I treat the three strategies in such detail is not so much to promote a diverse indicator mix. (There are enough differences among the models to justify tracking all three, and you can do so in a couple of minutes each week.) The real task of this inquiry is to breathe new life into once powerful indicators that in time fell short. Thanks to adaptive trend-following logic the SVI and the T-bill/Dividend ratio have reclaimed their standing as credible investment strategies. As one who helped beat the drum for early versions of these indicators, the turnabout is gratifying.
The U.S. needs submarines — and General Dynamics (GD) happens to be very good at making them; ...
Abercrombie & Fitch Co. (ANF) has been on a five-year slide as the company had fallen out of fav...
Based in Broomfield, Colorado, and founded in 1997, Vail Resorts, Inc. (MTN) operates mountain resor...