TESTING THE INFORMATIVENESS OF NON-PRICE VARIABLES WITH MIDAS TOUCH

In this investigation, we retest the informational content of non-price variables, namely, open interest and trading volume measures. Through a detailed review of similar studies carried out in the past, we find a gap in options non-price literature with regard to the methodology applied and timeframe during which these studies are carried out. A  study of this nature requires a methodology that handles the mixed frequency data structure  well, and which uses the data for the period where the options market is more mature. Our study addresses these two requirements with the mixed data sampling (MIDAS) estimation method, which uses every observation in the higher frequency space to regress them with lower frequency variable, which  keeps the model  flexible and parsimonious. We also use the current decade data (2011–2016) extracted from the NSE website and CMIE-Prowess database for 10 sample stocks and Nifty50 index. We apply Bhuyan and Chaudhury’s (2005) approach to derive open interest and volume-weighted price estimates. The inference on the informativeness of non-price variables has been drawn on the basis of the results of the MIDAS estimation, in-sample fit and out-of-sample performance of the model. The MIDAS estimation results suggest that non-price variables have a stronger impact and can be useful in predicting the underlying asset price on the expiration day of options contract, because at least one of the polynomial distributed lags (PDLs) of non-price variables turns out to be statistically significant for our sample assets. The in-sample fit and out-of-sample forecast evaluation confirms the higher forecast efficiency measured through RMSE, MAPE and SMAPE, accuracy and quality gauged via Theil’s U1 and U2 coefficients, respectively, in favour of the mixed data sampling (MIDAS) model. These results and the results of the Wilcoxon test on squared predictive errors highlight the superior predictive power of open interest and volume-based MIDAS in comparison to the time aggregation and the distributed lag approaches with the same variables.