Abstract
In many empirical settings, researchers encounter data where a covariate is only partially observed through ranks, such as a "Top 30" list. Standard regression techniques often apply "identity scores" (simply using the rank 1, 2, 3...) which can lead to biased estimates and poor model fit. This paper introduces D-rank, a new class of score functions based on the moments of order statistics from a pre-decided distribution. We demonstrate that D-rank is asymptotically optimal for maximizing the correlation between a response and a score when the chosen distribution matches the location-scale family of the true covariate. Applying this method to investor attention data from a major Korean internet portal, we find statistically significant evidence that message board activity predicts next-day stock returns.
Introduction
In the era of big data, we frequently encounter "Top N" lists—the most popular songs, the most-read news articles, or the most-discussed stocks. While these lists provide a clear ranking, they hide the underlying magnitude of the variable they measure. For example, the difference in discussion volume between the #1 and #2 ranked stocks on a message board might be vastly larger than the difference between the #29 and #30 stocks.
When researchers attempt to use these ranks as predictors in a regression model, they face a challenge: how to transform a rank (an ordinal value) into a score (a cardinal value) that accurately reflects the underlying phenomenon. This paper addresses this challenge by proposing a distribution-guided scoring system that accounts for the likely distribution of the unobserved data.
The Problem with Identity Scores
The most common approach to ranked data is the use of identity scores, where rank $r$ is treated as the value $r$ itself. This assumes a linear relationship between ranks, implying that the distance between every consecutive rank is identical. However, most real-world popularity data follows heavy-tailed distributions, such as the Power-law or Gamma distribution. In these cases, the "winner" often commands a disproportionate share of the total volume. Using identity scores in these scenarios ignores the "heavy-tail" reality, leading to attenuated correlation estimates and potentially misleading statistical inferences.
The D-Rank Methodology
The core innovation of this work is the D-rank score function. Instead of using the raw rank $r$, we transform it into a score $S_n(r)$ defined by:
$$S_n(r) = E[Z_{(r:n)}]$$
where $Z_{(r:n)}$ is the $r$-th order statistic of a pre-decided random variable $Z$. Essentially, the researcher chooses a distribution (e.g., Normal, Gamma, or Power-law) that they believe represents the underlying covariate. The score for rank $r$ is the expected value of the $r$-th largest observation if one were to draw $n$ samples from that distribution.
Theoretical Advantages
- Asymptotic Optimality: The authors prove that if the chosen distribution $Z$ belongs to the same location-scale family as the true (but unobserved) covariate $X$, the D-rank scores are optimal for maximizing the correlation between the response and the predictor.
- Consistency: The least-squares estimator using D-ranks is shown to be a consistent estimator of the true correlation coefficient ($\rho$).
- Inference: The estimator asymptotically approaches a normal distribution, allowing researchers to perform standard hypothesis testing and derive $p$-values.
Diagnostic and Model Selection
How does a researcher know which distribution to choose? The paper proposes a diagnostic procedure based on the regression intercept. In a standardized model where the score function is correctly specified, the regression line should theoretically pass through the origin. If the estimated intercept deviates significantly from zero, it suggests that the assumed distribution for the ranks (e.g., assuming a Normal distribution when the data is Power-law) is misspecified. This allows for a data-driven approach to selecting the most appropriate score function.
Empirical Application: Investor Attention
To demonstrate the utility of D-rank, the authors analyzed a hand-collected dataset from Daum.net, one of Korea's largest internet portals. The data tracks the "Top 30" most discussed stocks on the portal's message boards over 537 trading days.
Findings
- Distribution Fit: The analysis revealed that internet discussion volume is best modeled by a Power-law distribution (specifically with $\alpha = 2.3$). This captures the "superstar effect" where the top few stocks receive an overwhelming majority of the attention.
- Predictive Power: Using the D-rank method, the study found a correlation of 0.043 between investor attention and next-day stock returns. While the magnitude appears small, the $p$-value was extremely significant ($p < 10^{-5}$), confirming that "buzz" on message boards is a statistically significant predictor of short-term price movements.
- Top-Rank Heterogeneity: The model also identified that the very highest ranks (1 and 2) sometimes exhibit non-monotonic behavior, suggesting that extreme levels of attention may introduce noise or reflect highly polarized investor sentiment.
Conclusion
The D-rank method provides a robust framework for regression analysis when covariates are only observed as ranks. By moving beyond simple identity scores and incorporating distributional knowledge, researchers can achieve more accurate correlation estimates and better understand the underlying dynamics of ranked phenomena. This approach is particularly valuable in digital economics and finance, where "Top N" lists are a primary source of information regarding consumer and investor behavior.
Summary updated