We are living through unprecedented times. Due to the ongoing global health pandemic, the international markets have plummeted with speeds never seen before, reminiscent of the 1930s and the Great Depression.
On February 19, 2020, the SP500 Index closed at an all-time high price and then proceeded to decline sharply over the following days, accumulating 7 consecutive daily losses. Then after a short rebound, it proceeded to go through dramatic circuit-breaking -9% and -10% daily losses, making the VIX volatility index reach all-time highs.
Now in this blog, we have discussed the VIX a handful of times already and in one of our earlier posts, we shared a Deep Learning algorithm to forecast it. It seems appropriate to revisit it to see if our algorithm could have forecast this peak.
In our implementation, our approach to forecasting VIX peaks is performed by feeding 30-day windows of VIX closing prices to a neural network (a ResNet for time series) and have it predict whether 20 days ahead the VIX will be above or below 15 points. We used data up to 20/10/2015 to train the ResNet along with synthetic scenarios created by a GAN trained on train data. The effect of adding synthetic scenarios in our experiment was quite remarkable, the loss curves became a lot smoother so the model became more robust and improved its accuracy significantly. Let’s see how it would have fared this year:
In this graph we can see all the predictions made, the last prediction that we can check is on 28/2/2020 because it’s the last day for which 20 subsequent trading days have already passed as of 30/03. At that point our ResNet had predicted the VIX to be above 15 points for 4 days straight already, starting at 25/2/2020 so it is evident that this forecast would have been key if used as a “panic signal” of an investment model.
On the rest of the test period, we can see that our ResNet believed the VIX to be calm up until 24/2/2020 but gave off some warning signs around 12/2/2020 although they were not as conclusive, as the prediction went back to “calm” on the following week.
The conclusion to this post is evident, forecasting a time series such as the VIX is extremely difficult, but this time around our algorithm was able to warn us of the dangers that were about to come.