When Career Data Runs Out: Structured Features and the Signal Limits of Founder Prediction

Can a resume predict which founders will succeed?

That is the question Yagiz Ihlamur's paper puts to the test. Take 4,500 real founder career histories, where only 9% went on to succeed, and try to pick the winners using nothing but what is on paper: where they worked, what they studied, what they built before. The result is a story in two acts. Careful engineering of that career data works far better than asking an AI model directly, more than tripling the precision of random guessing. And then it stops working, for a reason that matters to everyone in founder prediction: 84% of the founders who succeeded look statistically identical to those who failed on every career metric the paper could measure. Whatever separated them, conviction, timing, the idea itself, never made it onto the resume.

What this paper contributes to founder prediction and quant VC

Most work on VCBench asks how far LLM reasoning can push founder success prediction. This paper asks the opposite question: how far can classical machine learning over carefully engineered structured features go, and where does the signal in career data run out? The answer is a structural baseline for quant VC. Twenty-eight features engineered directly from raw JSON career fields, combined with one deterministic rule and XGBoost boosted stumps, reach a validation F0.5 of 30.3%, a +17.7 percentage point improvement over a zero-shot LLM baseline.

The deeper contribution is a characterization of the information ceiling of career data alone. Across 120+ experiments, every model family converges at a cross-validation F0.5 of roughly 0.25 and a validation F0.5 of roughly 0.30. The paper argues this is a dataset ceiling, not a modeling ceiling, and identifies exactly what a richer benchmark would need to include to push past it. That makes the work function as a benchmark diagnostic for VCBench itself.

How the system works as a quant VC baseline

The public VCBench dataset contains 4,500 founder profiles with a 9% success rate. Each profile carries two representations: an anonymized prose paragraph summarizing the career, and the raw JSON fields that prose was generated from (jobs, educations, IPOs, acquisitions). Most prior work operates on the prose because it is immediately readable by LLMs. This paper parses the JSON directly, recovering exact company size buckets, role seniority codes, and exit counts, quantities that survive poorly in prose renderings.

A four-tier structured feature set. Tier 1 captures exit signals: founders with one prior exit succeed at 22.8% in training data, versus an 8.5% base rate for no-exit founders. Tier 2 captures sacrifice signals, operationalizing what a founder gave up to start the company, such as the prestige gap between the largest pre-founding employer and the founding itself. Tier 3 crosses education with relevance: false-positive analysis shows misclassified failures have higher average education prestige than true positives, so what matters is whether the degree is relevant to the founding domain. Tier 4 captures career trajectory shape: seniority monotonicity, company size trajectory, industry pivot count, and serial-founder interaction terms. All 28 kept features have zero null rate.

One deterministic rule beats three. Before the classifier, a single rule fires: if the founder has any prior exit, predict success. On training data this rule alone reaches 24.5% precision, 2.7x the base rate. Two additional candidate rules (top-10 school plus STEM plus founder role, and C-level plus serial founder) were tested and dropped after validation analysis showed they amplified false positives.

Decision stumps, not deep trees. Every model family tested, XGBoost, LightGBM, RandomForest, LogisticRegression, and stacking ensembles, converged on the same answer: trees of depth one. With only 405 positives, deeper decision boundaries cannot be estimated reliably. The final configuration is XGBoost with max_depth = 1, selected by 200-trial Bayesian hyperparameter optimization, with scale_pos_weight = 10 to upweight the 9% positive class.

How accurate is the structured baseline?

Headline numbers on VCBench:

• Validation F0.5: 30.3%, precision 33.3%, recall 22.2%.
• +17.7 percentage points over the zero-shot LLM baseline (F0.5 of 12.7%).
• Private test set (4,500 held-out rows, three folds): F0.5 of 28.1%, precision 32.8%, a 3.6x lift over the 9% base rate, consistent with the validation ceiling.

Position on the VCBench leaderboard at the time of writing:

• Random Rule Forest: F0.5 of 28.1%.
• This paper (structured features + rule layer): 30.3%.
• Policy Induction: 34.0%.
• Think-Reason-Learn / Verifiable-RL: 36.6%, the top entry.

The approach does not reach the top of the leaderboard, and the paper is direct about that. What it establishes is what direct JSON parsing with classical ML achieves on this benchmark, and the ceiling it hits.

The LLM feature experiment: prose adds nothing to the JSON

The paper runs a controlled experiment that matters for every LLM-augmented quant VC pipeline: extract nine features from the prose field using Claude Haiku (domain expertise depth, conviction indicators, career narrative type, highest seniority reached, prior founding attempts), and add them to the structured model. At 67% dataset coverage the LLM features looked promising, improving cross-validation F0.5 by +0.73 points. At 100% coverage the effect vanished: a delta of -0.05 points, a null result within one standard deviation.

The failure is structural, not a prompting problem. The model allocates 26.4% of its importance budget to the LLM features, but this comes directly out of the structured features' budget: it redistributes, it does not add. The reason is that VCBench's prose is generated from the same JSON fields the structured model parses directly. The prose is a lossy re-encoding, not a richer source: ordinal company-size buckets collapse into imprecise language, null signals are silently omitted, and field interaction terms do not survive text rendering. Extracting LLM features from the prose version of the same JSON is doubling down on the same signal, not adding a second source. The paper notes the caveat explicitly: in datasets where prose is written by humans, such as founder bios and pitch decks, LLM extraction may surface signals genuinely absent from structured fields.

The information ceiling and the two-population structure

The core finding: 84% of successful founders in this dataset have no prior exits, and they are, by every structured metric the paper measured, statistically indistinguishable from failures. The rule layer captures the exit-founder minority at high precision. Among the non-exit majority, the best single separator found (industry alignment) carries a 1.38x success-rate ratio, weak enough that it contributes noise as often as signal. All 120+ experiments confirm the same structure: once exit founders are handled by the rule layer, no combination of structured features moves the cross-validation needle materially for the rest.

Pushing past the ceiling requires signals that do not exist in this dataset: idea quality, market timing, team composition, investor network effects, and macroeconomic context at founding. This is directly useful information for the benchmark's evolution. A benchmark where the best-engineered structured features and the best LLM extraction from prose converge on the same answer is a benchmark that has characterized its own information content, and identified exactly where the next version should push.

Where this fits in the quant VC research landscape

This is independent research by Yagiz Ihlamur (Amazon), built on VCBench, Vela's public quant VC benchmark. It complements Vela's own threads from the opposite direction. Where the Think-Reason-Learn family (GPTree, Random Rule Forest, Reasoned Rule Mining, Policy Induction) uses LLMs to emit decision logic, and the LLM-Augmented ML thread (LLM-AR, GPT-HTree, Rare-Event Prediction, Verifiable Reasoning, Learning What to Ask) puts LLMs inside classical pipelines, this paper establishes what classical ML achieves with no LLM in the loop at all, and proves via ablation that on this benchmark, LLM feature extraction from generated prose is redundant with direct JSON parsing.

The result strengthens the case for both threads: the gap between this structural baseline (30.3%) and the leaderboard top (36.6%) is a measure of what LLM reasoning genuinely adds beyond structured features on career data.

Limitations

The paper is explicit about three caveats. The information ceiling characterization is dataset-specific: different labeling criteria, time horizons, or data collection pipelines may yield different ceilings. The LLM redundancy result holds for prose generated by the VCBench pipeline; human-written prose such as founder bios or pitch decks may carry signals genuinely absent from structured fields. And the private test set is held by the contest organizers and was not independently replicated, though the fold-level results provide robustness evidence.

Read the paper

When Career Data Runs Out: Structured Feature Engineering and Signal Limits for Founder Success Prediction.
Yagiz Ihlamur. IEEE, 2026.

• IEEE Xplore: ieeexplore.ieee.org/document/11638128.
• arXiv: arxiv.org/abs/2604.00339.
• Code: github.com/ihlamury/vcbench.

The paper is benchmarked on VCBench, Vela's public quant VC benchmark. For the methods it compares against, see Policy Induction, Random Rule Forest, and Think-Reason-Learn.

For research collaboration on quant VC, founder success prediction, or benchmark design, email engage@vela.partners.