I just wanted to share with you the results from my next generation S&P500 Volatility Model. This model seeks to predict, on a weekly basis, the direction of the S&P500′s Historical Volatility. I optimize the model weekly and then make a one week forward prediction on a rolling time series window.

To test how robust the model is, I took a sample from the training data to see if the model would be able to match the actual trend. As I said previously, this volatility model has a predictive accuracy between 60 to 70%, in this case its 65%.

I highly doubt that I’d be able to do better than 70% in predicting volatility, because of its “volatile nature,” but as a future option trader I’m willing to live with it. :)

File: S&P500 Volatility Prediction Results

Is that it’s too damn volatile! I’ve been working on the next gen version of a directional volatility model and its been both frustrating and rewarding at the same time. As long as the model is re-optimized on a weekly (or daily) basis, the predictions tend to have a 60-70% accuracy.

The problem is that you can’t forward forecast it 1 year into the future because volatility is, well, “unpredictable!” :)

Gasp, I said unpredictable (Christian please feel free to jump in)! Yes, its true volatility is unpredictable far out in the future BUT in smaller time frames the “stickiness law” works nicely. Noble prize winner, Robert Engle, proved (link coming) that volatility tends to cluster and that times of low volatility seem to be followed by more periods of low volatility. The same “pattern” exists for times of high volatility and I call this the stickiness law.

So predicting the direction of volatility is feasible, based on my findings, as long as you continually update your model.  Since we just saw a period of high volatility recently, we’re likely to see additional periods of high volatility soon. Guess that makes sense since September and October is coming!

I haven’t been posting an option trades from my virtual account lately because I haven’t made any. The reason why is because forward test results indicate the model is more lucky than robust. So I’ve developed a second generation model that’s a lot better statistically and uses pattern recognition learners from YALE/Rapidminer (the official Rapidminer 4.0 version is out now). Initial results have been very good; in a forward test from Jan 1, 2007 through July 27th, the model issued 21 correct signals out of 31 total (roughly 68%).

What’s next for this model? Well I can’t tell you very much because its super secret but if things work out, I’ll be making an announcement in September.

I had a way better results in my virtual options account this week. My S&P500 short straddle worked out nicely and my overall account gained +0.5%. However, the QQQQ volatility model seems to be wrong more than its right so I’ll have think about how to tighten that up.