Network Contagion in Financial Labor Markets: Predicting Turnover in Hong Kong

Abdulla AlKetbi; Gautier Marti; Patrick Yam; Raed Jaradat

REVIEW 4 major objections 5 minor 14 references

Network Contagion in Financial Labor Markets: Predicting Turnover in Hong Kong

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Financial professionals are 23% more likely to leave their firm once 30% of their coworkers have quit within six months, and network-derived signals lift turnover prediction by about 30% over baseline models.

desk verdict A credible walk-forward turnover prediction result on a new regulator-sourced dataset, with a headline 'contagion' claim that is really an uncontrolled association. read the letter →

arxiv 2509.08001 v1 pith:X7EBAFXK submitted 2025-09-07 cs.SI cs.LGstat.AP

Abdulla AlKetbi , Patrick Yam , Gautier Marti , Raed Jaradat This is my paper

classification cs.SIcs.LGstat.AP

keywords employeeturnovernetworkcontagiontemporalnetworksfeaturepropagationHongKongSFCregisterpeereffectspredictionworkforceanalytics

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that quitting in Hong Kong's financial sector spreads through professional networks — that turnover is, in part, contagious — and that the contagion can be measured and used for prediction. Using the public register of licensed professionals kept by the Hong Kong Securities and Futures Commission, the authors build month-by-month co-employment networks covering 121,883 professionals and 4,979 firms from 2007 to 2024, then propagate turnover-related features across those networks. They report that a professional whose peers have departed en masse — over 30% within six months — is 23% more likely to leave, and that network-propagated features lift a gradient-boosting model's average precision by 30.2% and AUC by 9.1% over a non-network baseline in strict walk-forward tests. If correct, the result gives regulators, firms, and policymakers a forward-looking measure of workforce instability derived from records that already exist, and it reframes turnover as chiefly an organizational and relational phenomenon: firm stability and peer effects account for over 80% of predictive importance, while name-inferred demographics contribute under 5%.

What carries the argument

The load-bearing object is the temporal co-employment network: professionals as nodes, edges for simultaneous employment at a firm, weights proportional to overlap; firms connected by shared employees. Across monthly snapshots, a graph-based feature propagation operator — the weighted average of neighbors' features, f'_i = (Σ w_ij f_j)/(Σ w_ij), optionally iterated — diffuses signals of peer departures, tenure, and firm stability through both graphs. This operator is what converts raw registry events into contagion measurements, and its outputs are what the models consume alongside individual and firm covariates. The evaluation design — walk-forward validation with a one-month gap, undersamp

What would settle it

Re-run the Section 6.3 threshold analysis while controlling for the firm's own average departure rate, or with firm fixed effects. If the 23% excess departure probability shrinks to zero once firm-level stress is held constant, peer departures carry no independent contagious signal — they simply mark an unstable organization. A second check: test whether the effect persists for professionals whose peers departed to clearly different firms, which would isolate peer influence from shared workplace shocks.

Watch

Extended reading notes

Core claim

The central claim: turnover in a regulated financial labor market spreads through co-employment networks, so a critical mass of peer departures raises a professional's own departure risk — specifically 23% above baseline once over 30% of coworkers leave within six months. The predictive claim: next-month turnover is forecastable in strict walk-forward tests, and network signals carry information that individual and firm covariates do not; a network-augmented LightGBM beats a non-network baseline by +30.2% average precision and +9.1% AUC across 169 monthly test periods. The structural claim: firm-level stability (43.6% of importance) and network-propagated features (39.6%) dominate individual

Load-bearing premise

The 23% contagion estimate assumes that seeing coworkers leave makes a professional more likely to leave for reasons beyond the fact that the whole firm may be in trouble — the analysis does not hold firm-level stability constant, even though firm stability is the model's single strongest predictor.

Editorial extensions

If this is right

A regulator holding such a register could score every licensed firm for imminent instability, using the aggregate departure probability of its professionals as an early-warning indicator.
Firms could identify teams at risk of cascading departures and intervene before the 30%-peer-departure threshold is crossed.
Because firm stability and network features dominate demographics, retention policy aimed at individual attributes is likely to underperform policy aimed at organizational stability and team cohesion.
Predictive gains concentrate in stress periods — accuracy peaks after the 2009 crisis and declines as the market matures — so network monitoring is most valuable in turbulent markets.
The framework transfers to other regulated labor markets with licensing registries, such as Singapore or London, though threshold values would need re-estimation locally.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The 'contagion' reading of the 23% estimate presumes peer departures push the focal professional out, rather than merely flagging a firm in distress; re-estimating with firm fixed effects or firm-level turnover controls — both constructible from this register — would separate the two.
The headline +30.2% AP gain pairs LightGBM-plus-network against Random-Forest-without-network; the paper's own same-model comparison (Random Forest both ways) isolates +20.7% AP from the network features alone, a useful calibration for anyone benchmarking the framework.
A direct extension: weight the propagation operator by recency or seniority (recent co-workers or more senior peers weighted more) and test whether the 23% threshold response strengthens — that would distinguish simple herding from hierarchical influence.
Because demographics were LLM-inferred from names rather than self-reported, the negligible 4.3% importance is a statement about name-based proxies; a registry with verified demographic attributes could yield a different answer.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Worth reading for the predictive result: the SFC register is a genuinely new and clean data source, and the walk-forward evaluation is done with discipline. 169 monthly snapshots, a one-month gap to avoid leakage, 5.5 million predictions, and a +30% AP gain over a no-network baseline—that is a real, honestly presented result. The graph feature propagation is just weighted neighbor averaging, but the application to a market-wide labor network is new, and the robustness checks around undersampling and calibration are more than most papers do.

The soft spot is exactly where the authors put their headline. The 23% "contagion" estimate in Section 6.3 is a raw conditional association. Peers are co-workers at the same firm, so a firm-level shock (restructuring, poor performance, regulatory action) would produce exactly this pattern: many departures at once, and the focal professional more likely to leave. The paper does not control for firm fixed effects or firm-year shocks, reports no confidence interval, and the 30% threshold looks post hoc. The authors' own feature importance numbers make the common-shock reading plausible: firm-level stability features dominate. So the causal language is not supported; as a predictive association it is fine, but as contagion it needs a lot more work.

Minor issues: the demographic inference relies on a self-cited LLM method that is not validated here, though demographics are only 4.3% of importance, so this is almost a side note. The decline in predictability over time is interesting but not deeply explored; that is a missed opportunity, not a flaw.

The paper is for anyone doing applied workforce analytics, regulator-sourced data mining, or network-based prediction. The dataset and evaluation alone justify a serious referee. The contagion claim needs re-estimation with firm fixed effects, pre-specified thresholds, and sensitivity analysis; the authors should also release code and processed data. I would send it to review, with a clear expectation that the causal claim gets rewritten or heavily qualified.

Referee Report

4 major / 5 minor

Summary. The paper uses the Hong Kong SFC public register (2003–2024) to build monthly co-employment and firm-sharing networks for 121,883 licensed professionals and 4,979 firms. It proposes a graph-based feature propagation method that iteratively averages neighbor features over these temporal networks, then feeds the resulting signals into LightGBM/Random Forest models with strict walk-forward evaluation. The paper reports two headline results: (1) a 23% relative increase in turnover probability when more than 30% of a professional's peers depart within six months, and (2) a network-augmented LightGBM that improves Average Precision by +30.2% and AUC by +9.1% over a non-network Random Forest baseline. It also reports that firm-level stability and network-propagated features dominate feature importance while LLM-inferred demographics contribute little.

Significance. If the causal contagion claim and the predictive improvement both held as stated, this would be a valuable contribution: it uses a comprehensive, regulator-sourced temporal network that is uncommon in turnover research, and it evaluates predictions out-of-sample over 169 monthly periods with a one-month gap, which is methodologically stronger than much of the existing attrition-prediction literature. The walk-forward design and the scale of the dataset are real strengths. However, the current manuscript does not yet establish the headline contagion effect because the 23% estimate is an unadjusted conditional association, and the headline predictive gain is computed against a baseline that differs in both model and features. These issues are fixable through additional analysis, so the central promise is credible but not yet substantiated.

major comments (4)

[§6.3] The 23% contagion estimate is a raw conditional association: it conditions only on the fraction of co-employee peers who departed in a six-month window. Since the employee–employee network is defined by simultaneous employment at the same firm (§4.1), peer departures are strongly correlated with firm-level shocks, restructuring, and managerial instability. The paper's own feature-importance analysis (§6.2) shows firm-level stability is the largest contributor, making the common-shock alternative especially plausible. To support the causal 'contagion' language used in the abstract and conclusion, the authors must add controls for firm-level turnover, firm fixed effects, time fixed effects, and ideally individual tenure/role; report confidence intervals; and discuss the residual identification problem. If such controls are infeasible, the claim should be presented as an association, not a
[Table 2] The headline +30.2% AP improvement compares LightGBM with network features to Random Forest without network features. This confounds model choice with the network-feature contribution. A proper ablation must compare the same base model with and without network features — e.g., LightGBM(no network) vs LightGBM(with network) — and report paired tests across the 169 test periods. Using the numbers shown, Random Forest with network vs Random Forest without network gives about +20.7% AP, which is smaller than the advertised +30.2%; the current headline therefore overstates the value of the network features themselves.
[§6.3, threshold selection] The 30% peer-departure threshold and six-month lookback window appear to be selected post hoc from the same data that is used to report the 23% effect. No confidence interval, sensitivity analysis, or out-of-sample replication is given for this statistic. Thresholds chosen on the test set are known to inflate apparent effects. Please provide a continuous dose–response curve, a sensitivity sweep over thresholds and windows, and bootstrap or hold-out confidence intervals for the reported relative risk.
[§3.3–§5.1 and Algorithm 1] The no-leakage guarantee is not fully documented for all 36 features. The walk-forward protocol is described, but it is not stated explicitly that every feature used at month t — including edge weights normalized by 'career length' (§4.1), individual mobility-rate features, and the propagated features of Algorithm 1 — is computed using only information available at time t. If any of these quantities uses post-t employment spells, the reported out-of-sample gains would be inflated. Please provide an explicit temporal-ordering statement and, ideally, a leakage-check ablation or code release.

minor comments (5)

[Abstract vs §3.1] The abstract says the register spans 2007–2024, while §3.1 and Table 1 describe the dataset as 2003–2024. Clarify the discrepancy.
[§6.4 / Figure 3] Figure 3 is described as AP/AUC/F1 of the random forest model, while the main model of the paper is LightGBM. Specify why the random forest is used for the temporal plot and whether LightGBM shows the same pattern.
[§6.3] The '23% relative increase' is not defined precisely. State the baseline probability and the absolute change, and report whether the statistic is measured on the test set or the full sample.
[Figures 1–3] Figures lack clear axis labels and units in the text; readability would be improved with explicit legends and axis titles.
[Global] Several headings show typographical artifacts (e.g., 'T emporal Dynamics') and the text has minor spacing issues; a careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: walk-forward prediction is self-contained; minor self-citations are not load-bearing

full rationale

The paper's central contributions—walk-forward turnover prediction and network-augmented features—are not circular. The target is a one-month-ahead departure indicator; network features are computed from lagged peer departures and firm-level attributes known at month t, so the comparison in Table 2 is an honest out-of-sample evaluation. The propagation operator in Eq. (1) is a weighted neighborhood average that does not use the focal node's own target, so it cannot encode the outcome by construction. The §6.3 contagion statistic is an unadjusted conditional association; while the 30% threshold and six-month window appear data-derived and no confidence interval is reported, this is a statistical robustness limitation, not a case of the result being defined as its input. The only self-citations are [1] (prior descriptive mapping of the same public register) and [2] (LLM-based demographic inference). Citation [2] supports a secondary demographic-importance claim, not the main predictive or contagion results, so it is not load-bearing. No uniqueness theorem or ansatz is imported to force the model choice. Overall, the derivation chain is self-contained and the headline predictive result stands independently.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The analysis rests on three families of assumptions: the register is a faithful point-in-time record of employment; co-employment edges capture meaningful peer influence; and observed peer-departure associations are not confounded by firm-level shocks. The demographic conclusion also depends on the accuracy of the authors' self-cited LLM name-inference method. These are structural assumptions, not derived results.

free parameters (5)

Peer-departure threshold = 30%
Section 6.3; chosen post hoc with no sensitivity analysis, defines the 23% effect.
Peer departures lookback window = 6 months
Section 6.3; window is fixed without reported sensitivity.
Feature propagation iterations k = 1
Section 5.1; authors state k=1 captures most signal, an ad hoc modeling choice.
Undersampling ratio = not reported
Section 5.2; majority-class undersampling ratio is tuned but not disclosed; robustness check reports only +/-0.002 AP.
Edge weight normalization = overlap normalized by career length
Section 4.1; exact formula not given, and edge definitions affect propagated features.

assumptions (5)

domain assumption SFC register employment spells accurately record true employment periods and termination dates.
Section 3.1; all temporal graph construction and turnover labels depend on this.
domain assumption Co-employment edges are a valid proxy for meaningful peer influence.
Section 4.1; the contagion interpretation requires that working at the same firm creates meaningful social ties.
domain assumption Peer-departure associations are not confounded by firm-level or market-level shocks.
Section 6.3; the 23% claim is reported without controlling for firm stability, which Section 6.2 identifies as the dominant feature.
domain assumption LLM-inferred gender and country of origin are accurate enough to support the demographics conclusion.
Section 3.3 and Table 3; relies on self-cited Reference [2] with no validation in this paper.
domain assumption Features are point-in-time with no future leakage.
Section 5.2 walk-forward design assumes one-month gap and current-month features; no code to verify.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Network Contagion in Financial Labor Markets: Predicting Turnover in Hong Kong." pith.science (2026). https://pith.science/paper/X7EBAFXK

@misc{pith2026250908001,
  author       = {Pith},
  title        = {Pith review of: Network Contagion in Financial Labor Markets: Predicting Turnover in Hong Kong},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X7EBAFXK}},
  note         = {Machine review of arXiv:2509.08001}
}

read the original abstract

Employee turnover is a critical challenge in financial markets, yet little is known about the role of professional networks in shaping career moves. Using the Hong Kong Securities and Futures Commission (SFC) public register (2007-2024), we construct temporal networks of 121,883 professionals and 4,979 firms to analyze and predict employee departures. We introduce a graph-based feature propagation framework that captures peer influence and organizational stability. Our analysis shows a contagion effect: professionals are 23% more likely to leave when over 30% of their peers depart within six months. Embedding these network signals into machine learning models improves turnover prediction by 30% over baselines. These results highlight the predictive power of temporal network effects in workforce dynamics, and demonstrate how network-based analytics can inform regulatory monitoring, talent management, and systemic risk assessment.

Figures

Figures reproduced from arXiv: 2509.08001 by the authors.

**Figure 2.** Comparison of feature importance rankings across models. Both highlight [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Smoothed average precision (AP), AUC, F1 Score of the random forest [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 13 canonical work pages

[1]

In: Complex Networks and Their Applications (Complex Networks

AlKetbi, A., Marti, G., AlNuaimi, K., Jaradat, R., Henschel, A.: Mapping hong kong’s financial ecosystem: A network analysis of the sfc’s licensed professionals and institutions. In: Complex Networks and Their Applications (Complex Networks

work page
[2]

AlNuaimi, K., Marti, G., Ravaut, M., AlKetbi, A., Henschel, A., Jaradat, R.: En- riching datasets with demographics through large language models: What’s in a name? arXiv preprint arXiv:2409.11491 (2024)

arXiv 2024
[3]

science 329(5996), 1194–1197 (2010)

Centola, D.: The spread of behavior in an online social network experiment. science 329(5996), 1194–1197 (2010)

work page 2010
[4]

Journal of Applied Communication Research36(1), 56–73 (2008)

Feeley, T.H., Hwang, J., Barnett, G.A.: Predicting employee turnover from friend- ship networks. Journal of Applied Communication Research36(1), 56–73 (2008)

work page 2008
[5]

Scientific Reports14(1), 9036 (2024)

Gad´ ar, L., Abonyi, J.: Explainable prediction of node labels in multilayer networks: a case study of turnover prediction in organizations. Scientific Reports14(1), 9036 (2024)

work page 2024
[6]

Physics reports519(3), 97–125 (2012)

Holme, P., Saram¨ aki, J.: Temporal networks. Physics reports519(3), 97–125 (2012)

work page 2012
[7]

https://www.hkma.gov.hk/eng/key-functions/international-financial-centre/ hong-kong-as-an-international-financial-centre/

Hong Kong Monetary Authority: Hong Kong as an International Financial Cen- tre. https://www.hkma.gov.hk/eng/key-functions/international-financial-centre/ hong-kong-as-an-international-financial-centre/. Accessed: 2024-08-18

work page 2024
[8]

World Scientific (2016) 12 Abdulla AlKetbi et al

Masuda, N., Lambiotte, R.: A guide to temporal networks. World Scientific (2016) 12 Abdulla AlKetbi et al

work page 2016

Show all 14 references

[9]

Physical review letters86(14), 3200 (2001)

Pastor-Satorras, R., Vespignani, A.: Epidemic spreading in scale-free networks. Physical review letters86(14), 3200 (2001)

2001
[10]

In: Proceedings of the International AAAI Conference on Web and Social Media, vol

Perrotta, D., Johnson, S.C., Theile, T., Grow, A., de Valk, H., Zagheni, E.: Open- ness to migrate internationally for a job: evidence from linkedin data in europe. In: Proceedings of the International AAAI Conference on Web and Social Media, vol. 16, pp. 759–769 (2022)

2022
[11]

In: International Conference on Social Informatics, pp

State, B., Rodriguez, M., Helbing, D., Zagheni, E.: Migration of professionals to the us: Evidence from linkedin data. In: International Conference on Social Informatics, pp. 531–543. Springer (2014)

2014
[12]

Organization Science26(4), 1177–1191 (2015)

Vardaman, J.M., Taylor, S.G., Allen, D.G., Gondo, M.B., Amis, J.M.: Translating intentions to behavior: The interaction of network structure and behavioral inten- tions in understanding employee turnover. Organization Science26(4), 1177–1191 (2015)

2015
[13]

Social Network Analysis and Mining13(1), 28 (2023)

Younis, S., Ahsan, A., Chatteur, F.M.: An employee retention model using orga- nizational network analysis for voluntary turnover. Social Network Analysis and Mining13(1), 28 (2023)

2023
[2024]

Accepted for presentation

(2024). Accepted for presentation

2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

In: Complex Networks and Their Applications (Complex Networks

AlKetbi, A., Marti, G., AlNuaimi, K., Jaradat, R., Henschel, A.: Mapping hong kong’s financial ecosystem: A network analysis of the sfc’s licensed professionals and institutions. In: Complex Networks and Their Applications (Complex Networks

work page

[2] [2]

AlNuaimi, K., Marti, G., Ravaut, M., AlKetbi, A., Henschel, A., Jaradat, R.: En- riching datasets with demographics through large language models: What’s in a name? arXiv preprint arXiv:2409.11491 (2024)

arXiv 2024

[3] [3]

science 329(5996), 1194–1197 (2010)

Centola, D.: The spread of behavior in an online social network experiment. science 329(5996), 1194–1197 (2010)

work page 2010

[4] [4]

Journal of Applied Communication Research36(1), 56–73 (2008)

Feeley, T.H., Hwang, J., Barnett, G.A.: Predicting employee turnover from friend- ship networks. Journal of Applied Communication Research36(1), 56–73 (2008)

work page 2008

[5] [5]

Scientific Reports14(1), 9036 (2024)

Gad´ ar, L., Abonyi, J.: Explainable prediction of node labels in multilayer networks: a case study of turnover prediction in organizations. Scientific Reports14(1), 9036 (2024)

work page 2024

[6] [6]

Physics reports519(3), 97–125 (2012)

Holme, P., Saram¨ aki, J.: Temporal networks. Physics reports519(3), 97–125 (2012)

work page 2012

[7] [7]

https://www.hkma.gov.hk/eng/key-functions/international-financial-centre/ hong-kong-as-an-international-financial-centre/

Hong Kong Monetary Authority: Hong Kong as an International Financial Cen- tre. https://www.hkma.gov.hk/eng/key-functions/international-financial-centre/ hong-kong-as-an-international-financial-centre/. Accessed: 2024-08-18

work page 2024

[8] [8]

World Scientific (2016) 12 Abdulla AlKetbi et al

Masuda, N., Lambiotte, R.: A guide to temporal networks. World Scientific (2016) 12 Abdulla AlKetbi et al

work page 2016