[KongchangAI]
· 2 min read· 1,240 words

Match Quality Revealed by Employment Survival: The Best Fit Between People and Firms

Match Quality Revealed by Employment Survival: The Best Fit Between People and Firms

A large-scale two-tower neural network learns employee–firm match quality from employment survival, revealing its significant effects on turnover, earnings, and firm performance.

The study proposes a new way to measure employee–firm match quality, treating match as a pairwise attribute. Using a two-tower neural network, it embeds employees' career histories and firms' characteristics and business descriptions into a shared space, defines match quality as cosine similarity, and estimates separation hazard from employment relationship survival on over 1 billion person-periods of data. It finds that a 0.1 increase in match quality reduces separation odds by 48%; about one quarter of match quality variance is pair-specific; better-matched employees earn more, and their firms grow faster, are more profitable, and produce more patents; and the model can predict employee mobility destinations and goodwill recorded in acquisitions.

Introduction: There Is No Best Employee, Only the Best-Fit Employee

"There is no best employee in the world—only the best employee for a particular firm." This view comes from a large-scale empirical study that develops a new way to measure employee–firm match quality, treating match as a pairwise attribute and learning it from the survival of employment relationships. At the core of the study is a two-tower neural network that embeds employees (based on career history and current position) and firms (based on characteristics and business descriptions) into a shared space, with match quality defined as the cosine similarity between the two vectors.

Research source screenshot

Methodology: Learning Match Quality from Employment Survival

Two-Tower Neural Network and the Match Quality Measure

The methodological foundation of the study is a two-tower neural network. The employee tower encodes an individual's career history (e.g., past positions, industries, tenure) and current position information into a vector; the firm tower encodes firm characteristics (e.g., size, industry, age) and business description text into another vector. Match quality is defined as the cosine similarity between these two vectors, ranging from -1 to 1, with higher values indicating a better match.

The key to this design is that the functional form of match quality is taken from the "skill weights" model of firm-specific human capital, while the sole estimation target—the separation hazard—comes from viewing job match as an "experience good." In other words, the model does not directly predict performance or wages; instead, it infers match quality from whether employment relationships persist (i.e., survive).

Firm-Specific Human Capital and the Skill Weights Model

Firm-specific human capital refers to the skills and knowledge an employee accumulates within a particular firm that would lose much of their value upon leaving that firm. Classic theory (e.g., Becker) distinguishes between general human capital and specific human capital: the former is transferable across firms, while the latter is tied to a specific organizational context. In the "skill weights" model, employee skills are represented as a set of dimensions, and each firm assigns different weights to each skill; match quality is the degree of alignment between the employee's skill vector and the firm's weight vector. This setup explains why the same employee can exhibit systematic differences in productivity and willingness to stay across different firms, and it provides the theoretical basis for this study's definition of match quality as vector similarity.

Separation Hazard

Separation hazard is a core concept in survival analysis, referring to the conditional probability that an employee still in an employment relationship at a given time will leave at the next time point. Under the "experience good" perspective, employees and firms can only observe match quality after actually working together, so the worse the match quality, the more likely the relationship is to end. Researchers do not directly observe match quality; instead, they use large-scale employment relationship survival data and exploit the inverse relationship between separation hazard and match quality to estimate model parameters. This approach avoids relying on variables such as performance ratings or wages that may contain measurement error, giving the match quality estimates a stronger behavioral foundation.

Large-Scale Data and Estimation

The study estimates on more than 1 billion person-periods of data, covering 101 million employees and 8.4 million firms. This scale enables the model to capture subtle patterns that traditional small-sample studies cannot detect.

Key Findings: How Match Quality Affects Employees and Firms

Match Quality and Separation Probability

The study finds that a 0.1 increase in match quality reduces the odds of separation (leaving) by 48%. This effect is highly significant both statistically and economically, indicating that match quality is an important predictor of employee retention.

Variance Decomposition of Match Quality

About one quarter of the variance in match quality is specific to the employee–firm pair, while employee-level, firm-level, and occupation-level measures are set to zero by construction. This means that traditional measures based only on employees or only on firms would ignore this pair-specific variation and thus underestimate the importance of match.

Variance Decomposition and Pair-Specific Variation

Variance decomposition is a statistical method that splits the total variation of a variable into different sources. In this study, match quality is modeled as the sum of four components: employee, firm, occupation, and employee–firm pair. Because the model sets the employee, firm, and occupation measures to zero by construction, the remaining portion of the total variance is the pair-specific interaction effect. About one quarter of the variance belongs to this pair component, meaning that explaining match using only employee ability or firm characteristics would miss a large amount of fit information that cannot be attributed to either side alone. Conceptually, this result is analogous to an interaction effect in statistics: two factors may each look ordinary on their own, but together they produce a unique effect.

Returns to Employees and Firms

Employees with better matches earn more, especially early in the relationship. For firms, firms whose employees have higher match quality grow faster, are more profitable, and obtain more and more valuable patents over the next five years, outperforming industry peers.

Predicting Mobility and Acquisitions

Because the measure can score pairs that never formed, it can also predict where employees will move. In addition, when the employees of an acquisition target match better with the acquirer, the acquirer records a larger share of the acquisition price as goodwill, reflecting the market's recognition of match value.

Goodwill

Goodwill is an accounting concept referring to the portion of the consideration paid by an acquirer that exceeds the fair value of the target's identifiable net assets, usually interpreted as the pricing of intangible value such as brand, customer relationships, and synergies. In this study, when the target firm's employees have higher match quality with the acquirer, the acquirer records a larger share of the acquisition price as goodwill. The significance of this finding is that although match quality is not directly observable, market participants (acquirer management and auditors) have already partially capitalized it in pricing, providing external validation of the economic value of match quality that is independent of the model estimates.

Conclusions and Implications

This study provides strong evidence that employee–firm match quality is an important dimension independent of employee and firm characteristics. Match quality not only affects turnover rates but is also closely related to employee earnings, firm growth, profitability, and innovation output. For managers and HR practitioners, this means that recruiting and talent allocation should focus more on pair fit rather than individual ability alone. For researchers, this survival-analysis-based match measure opens new avenues for human capital research.

Research source screenshot

One-line Summary

A large-scale two-tower neural network learns employee–firm match quality from employment survival, revealing its significant effects on turnover, earnings, and firm performance.

Paragraph Summary

The study proposes a new way to measure employee–firm match quality, treating match as a pairwise attribute. Using a two-tower neural network, it embeds employees' career histories and firms' characteristics and business descriptions into a shared space, defines match quality as cosine similarity, and estimates separation hazard from employment relationship survival on over 1 billion person-periods of data. It finds that a 0.1 increase in match quality reduces separation odds by 48%; about one quarter of match quality variance is pair-specific; better-matched employees earn more, and their firms grow faster, are more profitable, and produce more patents; and the model can predict employee mobility destinations and goodwill recorded in acquisitions.

#employee-firm match quality#two-tower neural network#separation hazard#firm-specific human capital#variance decomposition#goodwill
Share:

Related articles