Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CrunchLLM: Multitask LLMs for Structured Business Reasoning and Outcome Prediction

Rabeya Tus Sadia, Qiang Cheng

TL;DR

CrunchLLM addresses startup exit prediction by leveraging Crunchbase's structured attributes and unstructured narratives via a domain-adapted LLM. It combines parameter-efficient fine-tuning (LoRA/QLoRA) with prompt optimization in a multitask setting to predict binary success and generate grounded explanations. The approach achieves state-of-the-art results on Crunchbase benchmarks, with accuracy exceeding $80\%$ and precision above $90\%$ for IPOs/acquisitions, and gains that are statistically significant over baselines ($p<0.01$). The introduced explanation traces enhance transparency for investors and policymakers, illustrating how funding, syndication breadth, and team size contribute to outcomes. Overall, CrunchLLM demonstrates how domain-aware fine-tuning and structured-unstructured data fusion can advance predictive modeling in venture finance.

Abstract

Predicting the success of start-up companies, defined as achieving an exit through acquisition or IPO, is a critical problem in entrepreneurship and innovation research. Datasets such as Crunchbase provide both structured information (e.g., funding rounds, industries, investor networks) and unstructured text (e.g., company descriptions), but effectively leveraging this heterogeneous data for prediction remains challenging. Traditional machine learning approaches often rely only on structured features and achieve moderate accuracy, while large language models (LLMs) offer rich reasoning abilities but struggle to adapt directly to domain-specific business data. We present \textbf{CrunchLLM}, a domain-adapted LLM framework for startup success prediction. CrunchLLM integrates structured company attributes with unstructured textual narratives and applies parameter-efficient fine-tuning strategies alongside prompt optimization to specialize foundation models for entrepreneurship data. Our approach achieves accuracy exceeding 80\% on Crunchbase startup success prediction, significantly outperforming traditional classifiers and baseline LLMs. Beyond predictive performance, CrunchLLM provides interpretable reasoning traces that justify its predictions, enhancing transparency and trustworthiness for financial and policy decision makers. This work demonstrates how adapting LLMs with domain-aware fine-tuning and structured--unstructured data fusion can advance predictive modeling of entrepreneurial outcomes. CrunchLLM contributes a methodological framework and a practical tool for data-driven decision making in venture capital and innovation policy.

CrunchLLM: Multitask LLMs for Structured Business Reasoning and Outcome Prediction

TL;DR

CrunchLLM addresses startup exit prediction by leveraging Crunchbase's structured attributes and unstructured narratives via a domain-adapted LLM. It combines parameter-efficient fine-tuning (LoRA/QLoRA) with prompt optimization in a multitask setting to predict binary success and generate grounded explanations. The approach achieves state-of-the-art results on Crunchbase benchmarks, with accuracy exceeding and precision above for IPOs/acquisitions, and gains that are statistically significant over baselines (). The introduced explanation traces enhance transparency for investors and policymakers, illustrating how funding, syndication breadth, and team size contribute to outcomes. Overall, CrunchLLM demonstrates how domain-aware fine-tuning and structured-unstructured data fusion can advance predictive modeling in venture finance.

Abstract

Predicting the success of start-up companies, defined as achieving an exit through acquisition or IPO, is a critical problem in entrepreneurship and innovation research. Datasets such as Crunchbase provide both structured information (e.g., funding rounds, industries, investor networks) and unstructured text (e.g., company descriptions), but effectively leveraging this heterogeneous data for prediction remains challenging. Traditional machine learning approaches often rely only on structured features and achieve moderate accuracy, while large language models (LLMs) offer rich reasoning abilities but struggle to adapt directly to domain-specific business data. We present \textbf{CrunchLLM}, a domain-adapted LLM framework for startup success prediction. CrunchLLM integrates structured company attributes with unstructured textual narratives and applies parameter-efficient fine-tuning strategies alongside prompt optimization to specialize foundation models for entrepreneurship data. Our approach achieves accuracy exceeding 80\% on Crunchbase startup success prediction, significantly outperforming traditional classifiers and baseline LLMs. Beyond predictive performance, CrunchLLM provides interpretable reasoning traces that justify its predictions, enhancing transparency and trustworthiness for financial and policy decision makers. This work demonstrates how adapting LLMs with domain-aware fine-tuning and structured--unstructured data fusion can advance predictive modeling of entrepreneurial outcomes. CrunchLLM contributes a methodological framework and a practical tool for data-driven decision making in venture capital and innovation policy.

Paper Structure

This paper contains 30 sections, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Large Language Models
  4. Fine-Tuning for Domain Adaptation
  5. Predictive Modeling of Startup Success
  6. Methodology
  7. Dataset
  8. Dataset Preprocessing
  9. Feature Engineering
  10. Data Challenges for Language Models
  11. Multitask finetuning
  12. Prompt Optimization for Supervised Fine-Tuning (SFT)
  13. Zero-shot inference
  14. Few-shot Supervision
  15. Implementation Details
  16. ...and 15 more sections

Figures (5)

  • Figure 1: Distribution of description length (in tokens).
  • Figure 2: CrunchLLM overall Workflow.
  • Figure 3: Prompt optimization ablation for classification and justification. Four prompt variants were explored from basic instruction (V1) to an optimized format (V4) to assess their impact on model performance. The optimized prompt (V4) showed improved alignment with expected output structure and interpretability.
  • Figure 4: Performance of prompt variants on classification and justification tasks. Optimized Prompt V4 achieves the best prediction scores, while justification quality remains stable, highlighting the impact of prompt optimization on multitask performance.
  • Figure 5: Effect of varying LoRA rank on accuracy. The best performance is achieved at rank-16, while both lower and higher ranks lead to reduced accuracy.