Human Capital in Context: What Employment Survival Reveals About Worker-Firm Match

Jung Ho Choi, Mario MilonearXiv 2026 · arXiv

TL;DR

The authors develop the Human Capital Transformer (HCT), a two-tower neural network that quantifies the "match quality" between specific workers and firms by learning from over a billion employment records. By treating fit as a property of the pair—rather than just the individual's skill—the model successfully predicts job durability, wage premiums, firm-level innovation, and the "goodwill" value in corporate acquisitions.

What Problem Does This Paper Address?

In traditional labor economics and HR, human capital is often treated as a scalar: a worker is "highly skilled" or "less skilled" based on education and experience. However, a brilliant research scientist might be a high-value asset at a biotech startup but a poor fit for a high-volume manufacturing firm.

The core challenge is that "match quality" is idiosyncratic and largely unobservable. Existing measures often focus on occupations (e.g., is this person a good fit for "Engineering"?) but ignore the specific firm (e.g., is this person a good fit for this specific company's culture and business model?). The authors aim to move beyond "good workers" and "good firms" to measure the value of the worker-firm pair.

Methods

The researchers introduce a two-tower neural network architecture to map both sides of the employment relationship into a shared $D$-dimensional embedding space.

  1. The Worker Tower: Uses a Transformer encoder to read a worker’s career history as a sequence of positions (similar to how a language model reads a sentence). It combines this history with static covariates (education, location) and the worker's current role to produce a vector $u_{wt}$.
  2. The Firm Tower: Uses a multilayer perceptron (MLP) to embed firm characteristics (industry, size) and a 384-dimensional sentence-transformer embedding of the firm’s "About" text into a vector $v_f$.
  3. The Metric: Match quality ($MQ$) is defined as the weighted cosine similarity between the worker and firm vectors. This aligns with the "skill-weights" model of human capital, where firms place different weights on various skill dimensions.
  4. The Objective: Unlike standard recommendation systems that try to predict the next item a user chooses, the HCT is trained on a separation hazard. It assumes that the "durability" of a relationship reveals its quality: good matches last longer, while poor matches dissolve. The model maximizes the likelihood of observed employment durations across 1.18 billion person-periods.

Datasets

The study leverages a massive scale of data:

  • Revelio Labs: Resume-level records for 101 million workers and 8.4 million firms in the US (2008–2024).
  • Compustat & CRSP: Financial data for publicly traded companies to link match quality to firm performance.
  • Kogan et al. (2017) Patent Data: Used to evaluate the relationship between workforce fit and innovation.
  • XBRL Filings: Used to extract "goodwill" and purchase price allocations for 1,334 corporate acquisitions.

Key Insights and Ideas

  • Fit is a Pair Property: A central finding is that 25-27% of the variance in match quality is specific to the worker-firm pair. This component cannot be explained by the worker being "generally good" or the firm being "generally a good employer."
  • The "Experience Good" Validation: Matching theory suggests that the value of a job is learned over time. The HCT confirms this: the association between match quality and wages is strongest in the first five years of a relationship and fades as the "surplus" is priced in or the worker becomes "locked in."
  • Counterfactual Capability: Because the model learns functions of characteristics rather than just fixed effects for observed pairs, it can score hypothetical pairings. It can predict how well a target company’s workforce would fit a potential acquirer before a merger happens.

Key Results

  • Predicting Turnover: A 0.1 increase in match quality (on a -1 to 1 scale) lowers the odds of a worker leaving the firm in a given year by 48%.
  • Wages: Workers earn approximately 7.9% more per 0.1 increase in match quality. Even when controlling for worker and firm fixed effects, a significant "pair-specific" wage premium remains.
  • Firm Performance: Firms with better-matched workforces grow faster. A 0.1 increase in aggregate firm match quality is associated with 5.4% faster revenue growth and 4.2% faster employment growth over the following year.
  • Innovation: Better-matched workforces are issued 16% more patents over a five-year horizon, and those patents are more valuable and more frequently cited.
  • M&A Goodwill: In acquisitions, if a target's workers fit the acquirer better than they fit the target, the acquirer records a larger share of the purchase price as "goodwill"—effectively putting a number on the value of the assembled workforce.

Limitations and Future Work

  • Associations vs. Causality: The authors emphasize that these results are associations. Because workers and firms choose each other based on information the model doesn't see, the coefficients describe what the measure tracks, not necessarily the causal effect of an exogenous change in fit.
  • The "Lock-in" Effect: The model relies on duration as a signal of quality. However, some relationships endure not because they are productive, but because moving is too costly (lock-in). While the wage and innovation data suggest the model is capturing productivity, disentangling fit from lock-in remains a challenge.
  • Temporal Generalization: The primary results focus on generalizing across different people. Future work could explore how these embeddings and match dynamics evolve over longer historical periods as industries and skill requirements shift.

Summary updated

Abstract

There is no best worker in general, only a best worker for a particular firm. We develop a measure of worker-firm match quality that treats fit as a property of the pair, learned from the survival of employment relationships. A two-tower neural network embeds workers, from their career histories and current positions, and firms, from their characteristics and business descriptions, into a common space, and match quality is the cosine similarity of the two vectors. The measure takes its functional form from the skill-weights model of firm-specific human capital, and its only estimated objective, a separation hazard, from the view of the job match as an experience good. Estimated on more than a billion person-periods covering 101 million workers and 8.4 million firms, a 0.1 increase in match quality lowers the odds of separation by 48%. About a quarter of the variance of match quality is specific to the worker-firm pair, which worker-, firm-, and occupation-level measures set to zero by construction. Workers earn more when better matched, especially early in a relationship, and firms whose workforces score higher grow faster, become more profitable, and are issued more, and more valuable, patents over the next five years than their industry peers. Because it scores pairs that never formed, the measure also predicts where movers go, and acquirers record a larger share of the price as goodwill when the target's workers would fit the acquirer better than they fit the target.
BibTeX
@misc{choi2026human,
  author = {Jung Ho Choi and Mario Milone},
  title = {{Human Capital in Context: What Employment Survival Reveals About Worker-Firm Match}},
  year = {2026},
  eprint = {2611.16058},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.16058},
}
Figure