Can global “macro” data be used to model local markets?
Correlation and linear models
A series of posts exploring the (lack of) linear relationship between USDINR, OIL and the NIFTY 50. Part I, Part II and Part III.
Using an SVM over dollar indices to trade the NIFTY
An SVM can be tuned with different parameters. Principal among those is the kernel to be used. Part I looks at the difference in predictive power when different kernels are used to train an SVM over the same dataset.
Part II tries to further tune the most promising kernel from Part I – the polynomial kernel. However, we find out that there is no silver bullet.
So, we take all the dollar indices (including USDINR) and observe the predictive power of an SVM using a polynomial kernel with varying degree parameters. We find that only two out of four indices are actually useful and that going too far back in history while training the SVM is counterproductive (Part III.)
Part IV trains an SVM using the two promising indices and their respective best-performing degree parameters from Part III. An ensemble model that chains the predictions looks promising.
While performing the experiments above, we noticed that none of these models side-step the 2018 drawdown. Their principal limitation is that they are “macro.” They will not handle local events well. So we combine the model from Part III with a Simple Moving Average and find that it leads to lower drawdowns in long-short portfolios. Part V also leads us to conclude that using just one of the indices is sufficient to meet our twin goals of shallower drawdowns and higher long-short returns.