A Data-Driven Look at VC Fundraising

Introduction

This project explores the venture capital (VC) fundraising landscape using a dataset of over 4,500 individual funds. The data includes detailed information on each fund’s characteristics—such as its age, investment stage, country focus, associated firm size, and more.

The Central Question:

Can we predict how much capital a VC fund will raise based on its fund structure, firm characteristics, and strategic focus?

Understanding which factors drive fundraising success is crucial for investors, founders, and policy-makers alike. Venture capital shapes the flow of innovation and access to opportunity—yet the fundraising process often seems opaque. This project aims to shed light on the patterns behind successful capital raises, using data-driven modeling to reveal which characteristics actually matter.

This analysis is particularly relevant to limited partners (LPs)—such as institutional investors, endowments, and high-net-worth individuals—who allocate capital across multiple venture firms. By examining how fund-level characteristics relate to fundraising outcomes, LPs can better assess a firm’s track record and compare it against industry-wide patterns. This model may support more informed decision-making when evaluating which VC firms demonstrate consistent fundraising success.

Dataset Overview

Relevant Columns for Prediction

Column Name Description
Fund Amount Raised Target variable — total capital raised by the fund (in millions)
AUM (Current) Assets under management for the firm managing the fund
Fund Age Age of the fund (in years)
Firm # of Funds Number of total funds the firm has launched
Average Fund Size (MM) Current average fund size at the firm
Fund Type Investment stage focus (e.g., Early Stage, Buyout, Mixed) — multi-label
Fund Country Focus Geographic target of investments (e.g., US, Europe)
Fund Status Whether the fund is currently raising, inevesting or closed
Fund Industry Focus Industry sectors the fund invests in (e.g., healthcare, tech) — multi-label

These columns were selected based on their potential predictive value and their availability prior to fundraising, to avoid data leakage.

Data Cleaning and Exploratory Data Analysis

Data Cleaning Steps

The raw dataset initially contained data on over 6,000 venture capital funds, which required substantial cleaning and transformation to ensure consistency and usability. Key steps included handling missing values, standardizing existing features, and creating derived variables from timestamps.

Preview of Cleaned Data

Fund Name Fund Status Fund Open Date Fund Type Fund Country Focus Fund Industry Focus AUM (Current) Firm # of Funds Average Fund Size (MM) Fund Amount Sought Fund Amount Raised Fund Age
01 Advisors 01 Fund Divesting 2019-05-03 Early Stage United States Other 855.00 3.0 285.00 200.00 135.00 6
01 Advisors 02 LP Investing 2021-01-20 Later Stage United States Internet Software/Services 855.00 3.0 285.00 325.00 325.00 4
01 Advisors 03 LP Investing 2022-03-24 Early Stage United States Packaged Software; Internet Software/Services 855.00 3.0 285.00 325.00 395.00 3
01fintech LP Investing 2022-01-01 Buyout United States Other 61.90 1.0 61.90 300.00 61.90 3
01vc Fund II LP Investing 2019-01-01 Early Stage China Other 14.57 3.0 14.57 14.57 14.57 6

Insights from Exploratory Data Analysis

Visual analysis of the dataset provided several key insights into the structure and characteristics of venture capital funds.

🌍 Fund Country Focus

Most funds in the dataset are based in the United States (2,377), followed by India with 213 funds. This trend aligns with expectations, as the U.S. has long been the epicenter of venture capital activity—particularly in regions like Silicon Valley and San Francisco.

🚀 Fund Type (Stage Focus)

The majority of funds in the dataset are focused on Early Stage investments. Some funds invest across multiple stages (e.g., Seed + Early + Late), but Late Stage-only funds are relatively rare. This reflects real-world dynamics: as investment rounds progress, the required check sizes grow substantially—often into the hundreds of millions or billions—making later-stage investing accessible to fewer firms.

📈 AUM vs. Number of Funds (Colored by Fund Age)

This scatter plot explores the relationship between the number of funds managed by a firm and its current assets under management (AUM). Each point represents a fund and is color-coded by its age.

There is a slightly positive trend: as the number of funds managed increases, AUM tends to rise as well. However, the relationship is weak and noisy, with wide variation in AUM even among firms managing the same number of funds.

Color gradients suggest that fund age is not a strong determinant of AUM. Older funds appear throughout the AUM spectrum, indicating that factors like investment strategy, fund type, or firm reputation may be more influential than time alone.

Grouped Table: Fund Type vs. Average AUM

This grouped table highlights the average assets under management (AUM) for funds operating at different investment stages. Funds that span multiple stages—especially those combining early stage, later stage, and buyout strategies—tend to manage significantly higher capital.

This suggests that broad-stage or hybrid investment approaches are associated with larger fund sizes. Rather than specializing solely in early or late stage, many of the highest-AUM fund types include a blend of strategies, which may signal greater flexibility, experience, or appeal to institutional investors.

Fund Type Average AUM (in Millions)
Early Stage; Fund of Funds; Secondary; Buyout 3,167.71
Seed Stage; Early Stage; Later Stage; Secondary… 2,102.77
Seed Stage; Early Stage; Later Stage; Fund of Funds 1,787.27
Early Stage; Later Stage; LBO; MBO; Buyout 1,777.40
Early Stage; Real Estate 1,764.08
… …
Early Stage; Mezzanine; Debt 28.42
Early Stage; Secondary 18.03
Seed Stage; Early Stage; Fund of Funds 11.46
Mezzanine; LBO; Real Estate 11.43
Seed Stage; Early Stage; Infrastructure/Proj Fin 4.81

This breakdown supports the broader finding that larger funds often diversify across multiple stages, likely due to the increased capital demands and longer investment horizons involved in managing a more flexible portfolio.

Imputation

For this project, I made selective decisions about how to handle missing data, balancing data quality with time constraints and model interpretability.

A notable variable, Fund Industry Focus, was excluded from modeling due to several challenges:

As a result, I opted to drop Fund Industry Focus entirely from the modeling dataset, instead of doing imputation. This decision was supported by the fact that most of the dataset (~80%) was missing industry data anyway, meaning the exclusion would not significantly affect model performance.

Framing a Prediction Problem

Prediction Type:

This is a regression problem, where the goal is to predict a continuous numerical value: the total capital raised by a VC fund.

Response Variable:

The response variable is Fund Amount Raised. This was chosen because it serves as a direct measure of a fund’s performance and credibility from the perspective of limited partners (LPs), such as institutional investors and endowments.

Although the dataset included a column called Fund Amount Sought, we chose not to use it for prediction. From the perspective of an LP making a funding decision, the amount a fund wants to raise is often not publicly known at the time of evaluation. It is fund-specific and can be aspirational, rather than predictive of actual outcomes. For this reason, it was excluded to avoid leakage and to align with what would be known at the “time of prediction.”

Why This Question Matters:

This prediction model can help LPs assess which fund characteristics are linked to stronger fundraising outcomes, potentially guiding their decisions when committing capital. It also contributes to a broader understanding of what types of VC funds tend to succeed in raising capital.

Evaluation Metric: R²

We use R² (coefficient of determination) as the primary evaluation metric for our regression model. R² measures the proportion of variance in the target variable (Fund Amount Raised) that is explained by the features in the model.

This metric was chosen over others because:

Time-of-Prediction Justification:

All predictor variables used in the model are features that would be available to a limited partner evaluating a fund at the time it is raising capital. These include fund-level descriptors like:

We intentionally excluded any features, like ‘Fund Amount Sought’ that are outcomes of the fundraising process or would only be known after the fact.

Baseline Model

Model Information:

The baseline model is a Lasso regression model trained on an initial set of features, without any non-linear transformations or advanced feature engineering. It was designed to serve as a simple benchmark, which I iterated upon in the final model.

Features Used:

Quantitative:

These were treated as continuous numerical variables and scaled using StandardScaler.

Nominal Categorical:

These categorical variables were one-hot encoded using OneHotEncoder(handle_unknown='ignore').

The preprocessing pipeline used a ColumnTransformer to scale numeric features and encode categorical ones:

preprocessor = ColumnTransformer(transformers=[
    ('cat', OneHotEncoder(handle_unknown='ignore'), ['Fund Status', 'Fund Country Focus', 'Fund Type']),
    ('num', StandardScaler(), numeric_cols2)
])

These steps were combined into a scikit-learn pipeline with a Lasso regressor (alpha=0.1):

pipeline = Pipeline(steps=[
    ('preprocessor', preprocessor),
    ('regressor', Lasso(alpha=0.1))
])

Evaluation Metric:

The model was evaluated using R², which measures the proportion of variance in Fund Amount Raised explained by the model. R² is intuitive, allows for direct comparison across models, and is standard for regression problems.

The R² values for the baseline model are quite close on both the training and test sets, which suggests that the model is not overfitting and has reasonably good generalization performance. However, this setup still uses relatively basic features and encodings.

Summary:

The baseline model establishes a strong starting point using interpretable fund-level features. It avoids data leakage by only using information available at the time of prediction. Although it does not yet capture complex relationships in the data, it sets a fair benchmark for evaluating the added value of the final model’s more advanced transformations.

To improve upon this baseline, the final model will:

Final Model

Additional Feature Engineering

In the final model, I added new features that better capture non-linear relationships and more meaningfully represent fund characteristics. These features were added because they reflect how fundraising works in the real world—for example, a fund’s age and its AUM are likely to influence how much capital it can raise, and not necessarily in a straight-line way. I included these based on common-sense patterns in venture capital.

New Features Added:

Model and Hyperparameter Tuning

The final model uses Lasso Regression, which reduces multicollinearity as well as unnecessary complexity. It performs automatic feature selection by shrinking irrelevant coefficients to zero. This improves interpretability and reduces overfitting.

Hyperparameters

I used GridSearchCV with 5-fold cross-validation on the training set to identify the best model configuration. My parameter search spanned:

param_grid = {
    'preprocessing__poly_feats__poly__degree': [1, 2, 3, 4, 5],
    'model__alpha': [0.5, 1, 3, 5, 10]
}

Best Hyperparameters found from GridSeach:

Preprocessing Summary

All feature transformations were combined into a single pipeline using scikit-learn’s ColumnTransformer. Polynomial features were generated for AUM (Current) and Fund Age, while all numeric variables were scaled. Categorical features (Fund Status, Fund Country Focus) were one-hot encoded, and Fund Type was multi-hot encoded using a custom transformer to reflect its multi-label structure. This setup ensured clean, consistent preprocessing during training and cross-validation.

Results

The final model achieved improved performance over the baseline model on both the training and test sets:

Model Train R² Test R²
Baseline 0.6991 0.7022
Final Model 0.7184 0.6684

While the improvement in Training R² is modest, it reflects meaningful gains in model fit. However, the testing R² decreases. This likely means that the model is overfitting to some data, and hence captures slightly less variance in Fund Amount Raised, likely due to the added nonlinear interactions and richer categorical encodings.

This improvement came without overfitting: the training and test R² values remain closely aligned. The final model does not generalizes well to the testing dataset, however the training R² benefits from regularization through Lasso, which helps reduce the impact of less informative features.

Conclusion

This project allowed me to explore the fundraising landscape of venture capital (VC) firms through the lens of data. By building a regression model to predict how much capital a VC fund will raise, I gained a deeper understanding of the many fund-level and firm-level characteristics that contribute to fundraising success—such as AUM, fund age, geographic focus, and investment stage strategies.

Beyond the modeling itself, this project introduced me to real-world venture dynamics and helped me appreciate how complex and multifaceted fund performance can be. It also deepened my understanding of key machine learning techniques, including feature engineering, regularization, and model evaluation, as well as how to think critically about the data generating process when designing predictive models.

Next Steps

I’d love to explore additional factors that could improve prediction accuracy or offer deeper insights into fund success, such as:

There’s a lot of potential to build richer, more nuanced models that don’t just predict outcomes but also inform strategy for both VC firms and limited partners.

References

Here are a few academic and industry sources that informed my understanding of VC fundraising and performance factors: