Monday, June 26, 2006

Technical Analysis Modeling

Been spending a lot of time recently on developing new technical analysis methods (commonly known also as "charting" - but the stuff I am developing barely involves a chart). In my academic career most of my research has involved applying time series models (statistical modeling methods related to regression analysis applied to data that is available as observations over time like historical temperatures, GDP, stock prices, population etc.). I have also done some work to apply these methods in the stock market.

I have found that, not surprisingly, it is pretty much impossible to forecast daily changes in stock prices using any standard time series model. This is why academic economists who work in finance say that technical analysis is rubbish and can't work. It is also why technical analysts use indicators which rather than forecasting changes in stock prices try generally to pick out turning points in the trend. I have created one indicator myself that is fairly useful by using an unusual combination of time series methods. My new approach is to try to forecast an indicator.

Recently, I have found that if you could predict the direction of the %K(5,5) full stochastic oscillator correctly (see this chart) you would beat the market by maybe 100% in bull market years and by hundreds of percent in bear market years. So being able to predict it is definitely a worthwhile thing. And it is far more predictable than stock prices themselves. Of course, no forecast can be 100% accurate and so these kinds of returns are not possible.

The time series model I have developed so far (in the last couple of days) can predict it well enough to increase the value of the account in bear markets. But in bull markets, following it blindly could lose money big time. Interestingly, we are now in a bear market by that definition. I tested each year from 1997-2006 and see how the model does over each year separately. This is called backtesting and is standard in developing technical analysis methods.

So I think my focus should be on improving those forecasts. Next step is to try some things that aren't in the time series textbook.

Anyway at the moment we have a very high probability forecast that the oscillator will be lower on Monday - i.e. stay short. This is backed up by the McClellan Oscillator and my E-Wave discussed in previous posts.

No comments: