May 2002
Technical Tweak
| "Consistency requires you to be as ignorant today as you were a year ago." |
| --Bernard Berenson |
ISITORS HAVE ASKED A LOT OF QUESTIONS about a lot of things, but the item getting the most inquiries is our technical indicators appearing on the Home Page. Many of you ask why their buy/sell readings don't change more frequently.
The short answer to the question is that they were designed not exhibit frequent changes. That, of course, begs for the longer answer.
First off, realize we're quants, not technicians. While we think technical indicators can be helpful in timing decisions, we certainly wouldn't put a lot of weight on them when analyzing individual securities much less portfolios (see Fundamental Technical Analysis in the Archives).
In an effort to incorporate technical analysis in our quant work, we devised an index to get a weekly reading of the markets as a whole. In July 1999, we explained the theory behind and how to interpret Our Technical Index.
Background
In short, our index is comprised of nine different technical indicators. These include advance/decline lines for the New York Stock Exchange and the Nasdaq, new highs and new lows on each index, as well as the relationship between the 200-day moving averages of two of our large cap portfolios.Each indicator can either be positive, negative, or neutral. Every Friday we get a weekly value for each indicator and then add them up to get a reading for the index.
Total values of +8 or +9 are buy signals while -8 and -9 are sell signals. Anything falling between is simply a neutral reading. Two buy signals without an intervening sell give the overall index a "buy". Similarly, two sell signals without an intervening buy give it a "sell" reading. The index moves to a "neutral" stance if a buy signal is followed by a sell or vice-versa.
One thing we noticed about many technical systems is that they quickly move from buy to sell and back again. Most investors would go broke from transaction costs if they attempted to act on each change. Even excluding costs, many technical systems would have you move in and out of the market too quickly to realize meaningful profits.
We didn't want our index to have this weakness so we gave it a relatively low sensitivity. Over our measurement period (more about this later), only 3% of the weekly index values fell within the buy or sell extremes.
So at least at one level, we can be satisfied that our index is doing what we hoped it would. But apparently, many of you feel it's too insensitive to changing market conditions.
Just consider, the last sell signal was back in September of 1998. A lot has happened since then including the dot-com blow-ups, last year's terrorist attacks, and most importantly, the first bear market in a decade. Maybe your complaints are right, at least it's worth considering.
Backtesting
As we understand it, the objection is not directed towards the actual composition of the index, but rather the interpretation of the data. That's good because it allows us to easily see how results would have differed had the index been more sensitive.|
Chart 1: Technical Buy and Sell Readings February 17, 1994 - February 28, 2002 |
 This chart compares the performance of the S&P 500 to the buy (green) and sell (red) signals from our technical index. Neutral readings are shown in grey. |
Data for the index extends back to the week ending February 17, 1994. There's nothing significant about this date, it's just when we started constructing our technical database. The actual technical index wasn't created until the launch of this site in July 1997.
Backtesting is a tool that's frequently used (and abused) in quantitative research. Essentially it's a way to see how a model would have reacted had it been used in the past. By plugging in historical data, you can see how the model's predictions would have compared to actual market results.
Since we had data back to 1994, we began by looking at how our existing technical model would have fared through the week ending March 1, 2002. Chart 1 shows the performance of the S&P 500 against its buy/sell/neutral readings.
If you took this approach, you wouldn't have been doing much selling. The model only flashed a sell signal for 9 of the 421 weeks of the measurement period. That's just over 2%.
On the other hand, it would have encouraged a lot of buying. Buy indications extended for 212 of the 421 weeks, fully 50% of the time.
One might think that this is a pretty bullish model given it would have had you buying from August of 1999 right though the bear market until the terrorist attacks in September 2001. From that point forward the index was neutral despite equities' sharp recovery.
Back-Up
In order to see if a more sensitive index would have been more valuable, we considered how the original could have been different. Again, we sought changes to the interpretation of the data, not the indicators themselves.The underlying idea behind the index is to bring together the disparate indicators in a way that can be easily interpreted. Since no single indicator is, in itself, more significant that the others a simple sum makes more sense than weighting some over others.
In fact, there are few changes you can make to the derivation of the index itself. It will always have a value from +9 to -9 so there are 19 different possible readings. As currently formulated, most are neutral with buy and sell readings only occurring with four (+8, +9, -8, -9). Obviously, sensitivity could be increased if more values were included in buy and sell readings.
But do you need more buy signals? As it is, it's already recommending buying 50% of the time. If investors really are risk-averse, does it make sense to be that aggressive? Wouldn't it be better to sell early rather than late? After all, if you sell early, you might regret not garnering the last few dollars of profit but if you sell late, you risk losing out on profits altogether.
So maybe its only the sell signs that need to be more sensitive. While there's no reason to weight one indicator over another, there's also no reason why buy and sell signals have to be symmetrical. What if the buy signals were left unchanged at +8 and +9, but sell signals were expanded to include -7 and -6?
|
Chart 2: Backtested Buy and Sell Readings February 17, 1994 - February 28, 2002 |
 Adjusting the sensitivity of the sell ratings yields more sell (red) recommendations and reduces the duration of the buy (green) signals. Accuracy is improved in 1996, 1997, and early 2000 although the model recommended buying into the bear market of 2001-2002. |
Using the same 421 weeks of data, weekly sell signals jumped to 27 from the current index's 9. Overall, the adjusted index would have had you selling in 45 of the 421 weeks. That's 11%, up from 3% in the original.
Over on the buy side, the adjusted index would have recommended buying in 147 weeks or 35%. That's 15% less than the original index.
But more important that the number of weeks the model would recommend buying or selling is what the market was doing when it made the recommendations. Chart 2 shows the adjusted index's buy and sell recommendations vs. the S&P 500.
The model would have had you selling through 1994's flat equity market. When the bull resumed in early 1995, it would have had you buying. When the rally stalled in mid-1996, it would have had you selling. Later that year, it flashed a buy signal that held until the Asian crisis of late 1998. The index remained neutral throughout for most of 1999, finally moving to buy in late July.
To its credit, the model would have had recommended selling as the market peaked in early 2000. On the other hand, it would have had you buying into the bear market from late 2000 to mid-2001. At that point it went neutral until recommending selling as the 4th quarter snap-back rally stalled.
Backtracking
No, the modified index isn't always right. It did miss most of 1999's run-up and did recommend buying at the beginning of the current bear market. Even so, it did flash well-timed sell signals in 1996 and 2000.Perhaps most importantly, as opposed to our original model, it did have sell signals. While they may at times be misguided, they're still more useful than the original index's ongoing buys.
Finally, you might be wondering if this entire process has been the result of data-mining. Data-mining is a major pitfall for any quant. Essentially it amounts to combing though data to come up with an explanation rather than using it to test a theory. Any time you rely on backtesting, data-mining is a distinct threat.
Nevertheless, in this case we'd plead not guilty. Recall that the indicators and resultant index are calculated in the same way that they always have been, it's just the way you read the index that we're altering. Yes, we considered how the model would have performed in the backtest, but it didn't determine the proposed changes.
It's not the past that's important anyway, it's how the model works now and in the future. Exchanging a little stability for accuracy seems like a good tradeoff. We'll see.
Search this site! Just enter you key word or words:
Get current quotes or follow your own custom portfolio,
courtesy of E-Line Financials:
|
|||||||||||
| Homepage | Return to Top |
