Here we outline the format of our current market and seasonal summaries.  For a more general view of the functionality of the site, check out our "spiel" page and a synopsis of our approaches to seasonal investing in general that was written for another investment service.

Below is a sample of a market summary…

In the leftmost cell, you see the date on which the indicators were current. Even with our most recent summary, all the data, with the exception of "annualized gain", lags by one day. Remember, the question we're asking is "which of yesterday's indicators best predicted today's big gains and losses?", with the aim of identifying trends.

Next, you see "average gain". The gain is annualized. If the value is negative, in fact, the whole cell will be gray, as opposed to white…the idea here is just to provide a bit of visual color-coding which might aid in the identification of patterns. Very roughly, the gain should parallel that of the Russell 3000.

What exactly do we mean by "annualized"?  We're simply projecting the gain or loss across a year time frame (about 252 trading days).  If we're looking at a full month (20 days) of historical data, and a stock gains 4% over that period, the gain will be 252*4/20, or 50.4%.  Looking at gains and losses in this way makes it easier to compare short term strategies against longer term strategies.

Next, there's "Market Focus". Here we take the highest % gain gotten from our indicators (you see this value at the top of the middle cell), and the lowest % gain (from the top of the right cell), and calculate how much these values deviate from a "chance" sampling of the data. Specifically, if "Market Focus" is 1.0, there's a difference of one standard deviation unit between the best indicator data and the best (NOT the average) random sampling. To make things simple, a high "Market Focus" means that the data is not noisy, and that the market over the measured period had a definite "direction" or "focus". If the "Market Focus" fell below zero (which almost never happens), that would mean you would probably have done better if you had thrown dice to make your stock picks.

In parentheses, we break down the significance into two chunks...the significance of our best strategy for gains, and of the best short strategy.  In the above table, you see that the top gaining strategy is quite significant (5.0), while the best short strategy is not tremendously significant (1.1).  Again, we rarely see negative values for either of these significances.

We may add more statistical indicators in the future...check here if you see any such indicators that are not explained on this page.

Next, you see "p=n=r=20, period = 180". Those are notes to ourselves.

Below that, you see that we took data for Feb 1999-2002 (four years of data), clumped it together in a single file, and analyzed it.

Then you move to the middle cell. It's also color-coded. The color relates to which indicator produced the biggest gains. It also prevents the page from being an ugly mess of numbers. You see "3monthgain: 48.81 to 1062.71(10)". This tells you that the strategy that produced the greatest gains was to buy stocks with a three month percentage gain between 48% and 1062% . The "(10)" tells you that this was the highest percentile or "slice" of data. In our attempts to find the most significant results, we may cut the data into 3, 5, 10, or 25 slices ( note). In the above example, it's easy to see that we decided to cut the data 10 times..look at all the values inside the parentheses and you'll see that the highest slice is "10". If you examine 3000 stocks, then 300 of them (give or take a couple) should fall into the 3 month gain range of 48% to 1062%.

For a list of all our indicators, click here.

To step backwards a bit, you see "135.7" before "3monthgain". That's the annualized gain that you'd get if you had bought the entire universe of stocks that had a 3 month gain between 48% and 1062% at the beginning of the trading session that followed the upperleft date.

The right cell follows the same logic as the middle cell, but here we highlight indicators that predicted losses. It's also color coded.

We're not finished. Take a look at the next summary…

Here, you see "ranked" data. To understand, we'll first explain a potential flaw of "unranked data". Let's say we get data from the past four Februaries, and clump the data together into a single file. Assume that in one of these Februaries, beaten-up stocks experienced very big gains, while beaten-up stocks only performed in a mediocre way in the other Februaries. Then we crunch the data and find that beaten-up stocks top our list of gain-producing indicators. Potentially, that thinking is a bit flawed…you might be led to believe that the absolute best strategy for February is to buy beaten-up stocks, when, in fact, that was only true (in a big way), for a single, anomalous February. The alternative is to rank the data for a single February (on a scale of 1-50, though it doesn't matter much), then clump that February's ranked data with the other February's ranked data. For any time period, each scale number (1,2,3...50) should appear with the same frequency as for numbers in any other time period. Without going into too much explanation, the end effect is that the results from one February don't "overpower" the results of the other Februaries.

When looking at ranked data, you won't see exact values for, say, the moving averages. You just see the average on a scale of 1-50. A "1" would mean that the average is low (the lowest 2% of all stocks).

A good portion of our historical data is ranked. We think it better reflects possible seasonal effects, as opposed to anomalies. Of course, when we examine only a particular period in a particular year, there's no "clumping" of data, and no need to rank it.

Look at the next summary…

Here, we've combined indicators. Specifically, we show that the best combination of indicators for a big gain is a cheap stock price and a high yearlong gain. There's no reason (other than processing power) we can't search for "triple indicators" or higher, but we've found that we often lose significance as we combine indicators. This result, in fact, makes us skeptical of neural nets and the like, where many indicators are examined in order to produce an output. If you can't combine two indicators to get a result superior to that of a single indicator, isn't it a bit unlikely that combining large numbers of indicators would be beneficial? Having said that, we do sometimes get superior results with combinations of indicators, and thus the combined results you see above.

We can always squeeze monster gains out of certain sectors of the market by using horrendously specific criteria (e.g. buy all stocks that sell between $4.22 and $4.89 AND have lost between 53 and 57% of their value over the last 3 months AND have extremely high volatility), but are such results really significant from a statistical point of view?   Often, they're not, so you won't see us trying to narrow the list of potential longs and shorts down to just a handful of stocks.

Finally, look below...

Look at those percentage gains.  A 2376% annualized gain for cheap stocks in the month of January?  No.  For every stock in our database, we've divided the annualized gain by the standard deviation of the stock in question.  This produces the big numbers, since our standard deviations are fractions.  For those in the know, the figure is simply the "Sharpe ratio" with the simplification that we set the risk free rate of return to zero.

What's the idea here?  It's a bit complex, but a simple explanation might go like this:  in a bull market, big winners tend to have high standard deviations, and laggards have low deviations.  In a bear market, big losers also have high standard deviations.  If we knew that the next few months or years were going to be bullish, we could simply buy stocks with high standard deviations, and profit above and beyond the market averages.

We, however, don't want to assume we can call the market.  Assume that last October was a big winner for stocks with high standard deviations.  What's more, it was an up month.  So now it's the beginning of October.  You decide to buy high deviation stocks, based on last October's data.  October turns out to be a losing month.  You lose your butt.  Your mistake:  you assumed that October would be a winning month.

In short, we try to minimize the effect of the market on our purchases by dividing gains by standard deviation.  This may seem like a lot of unnecessary and obfuscatory work, but we'd refer you to a test of the risk adjusted approach versus the non-risk adjusted approach.  You might be surprised at the gains via one, and the less-than-impressive results of the other.

We should point out that our "risk-adjustment" isn't just for conservative investors.  Nor should we imply that you can't suffer big losses with a risk-adjusted approach.  However, if you can't forecast the future of the general market, we think it's the way to go, and a good chunk of our historical data takes this form.  On the other hand, if you believe that you can forecast the market, you'll probably want to study our non-risk-adjusted data.

Risk-adjusted data may also have special appeal to option investors.  In some periods we've measured, volatile stocks have historically underperformed even in positive markets.  There are also periods where non-volatile stocks gain quite a bit more than one would ordinarily expect.  Our risk-adjusted data reveals these trends.

We haven't exhausted potential approaches to analyzing this historical data.  For example, we've experimented with ranking the outputs (the gains and losses, in red), as well as the values of the indicators.  This approach is something of a test of consistency over the years, and also reduces the data-skewing effect of giant gains or losses in a handful of stocks.  It doesn't, however, leave you with much of a feeling for exactly how much money you'd make over time by following a particular strategy.  

Another approach we've experimented with is to calculate the total percentage gain you'd have achieved by following a particular strategy over the years, starting from the oldest historical period available and working forward.  In theory, this could be quite a bit different than looking for the highest average gains that a particular strategy has resulted in over the years.  As an extreme example, suppose that one strategy results in 100% gains every year for 4 years, and a 100% loss on the fifth year.  Your average gain per year is a sweet 60% (100+100+100+100 -100 divided by 5), but you've lost all your money!  A better approach would be to multiply the gains for each year (since 100% is a doubling of portfolio value, the above result would be 2*2*2*2*0 = 0, which clearly reflects the fact that that particular strategy would have left you homeless).  In the real world, however, we've found that this approach and our usual approach give very similar results.

All the data above has been edited for the sake of demonstration…it's partially fictitious. Don't use it for predictive purposes.

*************************************

Footnotes:

From a strict statistical point of view, this "searching" for significance is a bit unkosher. We know that. Practically speaking though, we only run through a couple of different slices of data, so it doesn't make much difference…it's not as if we experiment with a myriad of different slices in an attempt to "get lucky".  The main point is just to cut out junky, meaningless data for the benefit of the user.

Seasonal Investments home page

Copyright © 2008 MarketSynopsis.com. All rights reserved.