How do we avoid overfitting in the Indian context

While calculating these single factor scores, it is quite easy (and tempting) to create scores in such a manner that back tested results align with the scores. In other words, portfolio of stocks with higher score outperforms portfolio of stocks with lower scores.

However, this would be a classic case of overfitting. Meaningful data for Indian companies is not available for more than 15 years. By meaningful, we mean enough data to back test “quintile” portfolios. In this context, we could have easily created scores such that higher quintile portfolios always outperform lower quintile portfolios.

We have however chosen to create these scores based on empirically proven research. In other words, are scores are derived from methodologies that have been studied on a much larger and broader data sets (longer history and multiple markets) rather than what would have worked on Indian stocks in the last 15 years.

A case in point being our value score. If we create 5 quintile portfolios from a universe of Nifty 500 stocks using the value score (detailed methodology for metric performance discussed here), we can clearly see that higher quintiles have not really outperformed lower quintile portfolios (see chart below) in the last 10 years. On the other hand, higher quintile portfolios have done remarkably well in the last 3 years. This is in line with the global trend of value picking up steam post the COVID-19 crash in March 2020.