Tag: correlation

This is the last part of the study. Part I, Part II

The reason why a linear model between NIFTY and USDINR built in Part II failed could have been because:

1. Weekly returns were not appropriate for the relationship. Perhaps INR affects NIFTY at a higher frequency.
2. There is no linear relationship because a rising/falling INR. Changes are not uniformly good/bad.

One way to visualize it is to plot the NIFTY returns density at different USDINR return thresholds. If there is no obvious difference in the densities between NIFTY returns when USDINR is positive vs. when it is negative, one could conclude that there is no straight forward relationship between the two.

Here is the NIFTY weekly returns density when USDINR is going up (the rupee is depreciating):

Note the curve when USDINR weekly returns are greater than 0.5% vs. when are greater than 2%. There is a bearish bias.

And, NIFTY weekly returns density when USDINR is going down (the rupee is appreciating):

If you juxtapose the above densities, it is apparent that when the rupee is appreciating, the densities skew right, And when the rupee is depreciating, there is a left skew. These charts show that there is “a” relationship – just not what can be captured by a linear model.

Code and density plots for NIFTY vs. OIL can be found on github.

This is a continuation of the correlation study of Part I
Our correlation study showed a -0.54 between NIFTY 50 and USDINR whereas a 0.21 with OIL. Here, we will use weekly returns of the NIFTY and USDINR to build a simple linear model.

Building a linear model

A weak correlation doesn’t usually lend itself to a useful linear model. To illustrate this point, have a look at the diagnostics below:

Ideally, the ‘Residuals vs. Fitted’ plot should show residuals evenly distributed around the zero line – it doesn’t. The Q-Q plot should lie on the diagonal – it is marred by heavy tails. Hence, we should scale-down our expectations from the model.

For this post, we will split the time-series that we have into a “training set” that goes from 2010-01-01 to 2015-12-31 and a “test set” that goes from 2016-01-01 to 2018-09-30. We will build the model with the former and test it with the latter.

Results

Predicted vs. actual weekly NIFTY 50 returns:

To test our model, we will give it the actual NIFTY 50 returns (x-axis) and plot the predict NIFTY 50 returns (y-axis.) The problem here is immediately apparent: it is heavily bullish! It consistently gives a positive prediction.

Long and Long-short cumulative returns:

If we use our model to go long-only (L) or long-short (LS), we get the cumulative returns shown above. The model is no better than buy-and-hold (at least it is no worse, so there is that.)

Take-away

A weak correlation between NIFTY 50 and USDINR is not much to work with and a linear model built over that relationship is no better than buy-and-hold. Given the narrative spun by the media, it is tough to wrap ones head around the results above.

We conclude with density charts of weekly NIFTY returns under different USDINR return thresholds in Part III.

Code and charts on github.