Humans are naturally biased, and nothing biases us like a winner. This is a straightforward explanation of growth bias. Because humans are growth-biased, stock markets are too. You don’t need science to observe growth bias—it’s conspicuous. However, you do need science to build models that can perform well despite growth bias, meaning models capable of predicting the evolution or decay of this bias.

Active investing struggles to beat the market because the science needed to overcome growth bias is novel, out-of-the-box, original, and falls within the domain of modern science. AlphaBlock’s [3N] methodology articulates this science of beating growth bias, despite its challenges. [1]

Building on the Systematic Cluster [2] and the Poor Design of MCAP [3], today let’s demystify this omnipresent growth bias.

The S&P 500 case study, with its 500 components and annualized returns plotted on a Cartesian graph versus relative percentile performance ranking, provides a statistical perspective in psychology. Even though psychology is implicit in the data and not an explicit variable, it influences outcomes.

Since the clustered data was fragmented, polynomial regression was utilized to map the data effectively, emphasizing the various curves and slopes present in the clusters.

Positive sloping relative performance scores

The analysis of clusters based on annualized return and relative percentile revealed several notable findings. Firstly, a consistent positive trend was evident across all years, suggesting that stocks with higher relative percentile scores generally exhibited superior performance. This observation underscores a winner's bias prevalent in each year's data.

Secondly, the highest decile occasionally exhibited intermittent upward shifts, indicating instances of remarkable market returns. Remarkably, these trends persisted across various starting points.

Upward sloping relative percentile performance top scores

The distinct positive slope in our data set not only confirms the presence of growth bias—indicating society's chase after winners—but also shows that global investments should be skewed towards companies delivering higher relative percentile performance. The higher the annual return of a company, the more likely it is to attract market participants and investments, thus skewing the top relative percentile rankers into a curve pointing higher.

Conversely, there is no such positive, upward-pointing curve at the bottom of the relative percentile rankings. There is less societal interest in the losers, leading to a lack of mass interest at the tail end, so much so that at times the polynomial regression lines point lower. The psychology is skewed against the losers (value bias) as much as it is skewed in favor of growth bias.

Downward sloping relative percentile bottom scores

If you are a loser, the market punishes you more, and vice versa. This is the classic tale of the Rich Get Richer and the Poor Get Poorer [4]. The composite picture reveals a chaotic map with boundaries.

Composite map of relative percentile performance scores

This is where the demystification of growth bias begins. First, growth bias does not operate alone. Second, growth bias does not imply certainty. Third, what constitutes growth bias today could transform into a no-bias or value bias tomorrow. Fourth, growth bias can be self-detrimental. The higher the slope and skew of the curve upwards, the more likely it is that the respective stocks will see profit-taking and possibly even trigger a momentum crash. Fifth, the polynomial regression lines indicate that growth bias, which produces extreme returns, does not perfectly intersect with large-cap, big-sized companies. Sixth, the beauty of the systematic cluster lies in its ability to map boundaries. Seventh, where there are boundaries, there is predictability. Eighth, relative percentile ranking lends itself to better training data. Ninth, better training data can lead to more effective machine learning systems.

So you see, growth bias, as illustrated by the systematic cluster, offers a straightforward way to build predictive systems. This current research is easily extendable to different time frames, and what works annually is bound to differ over three years, five years, or even less than a year.

Oliviu Cigan and Mukul Pal