lia-prop13/startup-Investments-analysis

📊 StartUp Investments EDA

---

1. Background & Objectives

This project explores a comprehensive dataset of startup investments (sourced from Crunchbase) to uncover the primary factors that predict a startup's survival and trajectory in a competitive market.

Through this Exploratory Data Analysis (EDA), we analyze historical funding data, investment rounds, and market categories to determine which variables drive specific company outcomes - namely, whether a business is ultimately "closed" or "acquired". In the context of this research, an acquisition serves as the most definitive indicator that a startup has survived and generated sufficient market value to be purchased and continue its operations.

---

2. Research Question

Based on historical investment data, what are the strongest predictors of a startup being acquired (an exit)?

• Rationale: By utilizing investment data to predict acquisitions, this analysis aims to identify the core indicators of startup survival. In this context, an acquisition serves as a strategic proxy: it signifies that a startup has generated sufficient market value and institutional validation to avoid collapse in a volatile environment.

---

3. Dataset Overview

Dataset Characteristics

The dataset provides a comprehensive look into the global startup ecosystem, utilizing data from Crunchbase—the premier platform for tracking private companies and investment trends. It encompasses extensive financial metrics, ranging from specific investment amounts across various rounds to debt and grants, alongside critical operational details such as market categories and company status. This specific version was sourced from Kaggle to analyze the factors driving startup survival and acquisition.

• Source: Access Raw Dataset Here

• Raw size: Approximately 54,000 rows & 39 features

• Clean size: 4,757 rows & 23 features

Target Variable and Sampling Strategy

The primary target for prediction is the company status, which has been transformed into a numeric binary format: 1 (Acquired) and 0 (Closed).

To ensure a clear and definitive predictive outcome, the dataset was strategically sampled by removing all "operating" records. This decision was driven by the inherent ambiguity of currently operating startups, which could either be on a path to success or on the verge of collapse. By focusing exclusively on finalized exit events (Acquisitions vs. Closures), we eliminate potential target leakage and establish a robust, balanced dataset of approximately 6,300 records for analysis.

• Class Distribution: The sampling process resulted in a relatively balanced distribution between the two target outcomes: 62% Acquired and 38% Closed. This ratio provides a reliable basis for analysis, ensuring that both categories are sufficiently represented to identify the key predictors of startup success versus closure.

Data Dictionary

The following table outlines the 23 features included in the final cleaned dataset, categorized by their role in the analysis:

Feature(s)	Category	Description
status	Target	The prediction goal: 1 for Acquired, 0 for Closed.
market, country_code, founded_year	Core Attributes	Descriptive indicators of the startup's industry, geography, and age.
funding_total_usd, funding_rounds	Financial Aggregates	Global metrics of total capital raised and the frequency of investment events.
round_A through round_H	Investment Stages	Capital raised specifically in each professional venture round.
seed, angel, equity_crowdfunding, product_crowdfunding	Early & Alternative	Early-stage funding sources often indicating initial market validation.
debt_financing, grant, private_equity, convertible_note, unclassified_funding	Specialized Funding	Various financial instruments used to sustain operations or drive growth.

---

4. Data Preprocessing & Cleaning

Initial Data Assessment

A diagnostic review of the raw dataset was performed to identify potential integrity issues and noise. The assessment revealed several critical areas requiring attention:

• Redundancy: Identification of 4,855 duplicate records within the raw data.

• Missing Value Density: Significant gaps identified in geographical features (e.g., state_code with >24k missing values) and founding dates.

• Data Gaps: A recurring pattern of missing information across approximately 4,856 financial records, indicating low-signal data points that could impact model reliability.

I. Structural Cleaning & Feature Selection

• Deduplication: Removed 4,855 duplicates to ensure unique entity representation.

  • (49,439 rows remain)

• Target Refinement: Isolated finalized outcomes (Acquired/Closed) by removing "Operating" and null records; prevents target leakage and ambiguity.

  • (6,295 rows remain)

• Dimensionality Reduction: Dropped 16 irrelevant columns, narrowing the focus to 23 high-impact features.

  Click to view the Feature Removal Log & Rationale

  To streamline the predictive model and eliminate noise, the following features were removed:

  • Identifiers & Web: name, homepage_url, permalink (Zero predictive signal).
  • Granular Location: city, region, state_code (Removed to prevent overfitting; country_code retained for generalizability).
  • Temporal Noise: Exact dates and months (e.g., founded_at, founded_month) removed to reduce noise; founded_year kept as a stable predictor.
  • Out-of-Scope Financials: post_ipo_equity, post_ipo_debt, secondary_market (Public/late-stage metrics outside research scope).
  • Ambiguous Data: undisclosed (Missing or hidden funding provides no actionable profile).
  • Redundancy: category_list (Streamlined in favor of the cleaner market feature).

II. Handling Missing Values (Imputation & Filtering)

• Temporal Data: Imputed founded_year (21.1% missing) using the median. This preserves data volume while providing a stable estimate for age-based calculations without introducing significant distribution bias.

• Identity Integrity: Removed records with missing market (3.67%) or country_code (9.98%). Since these features are core identity markers, imputing them would introduce synthetic noise and reduce the model's real-world reliability.

  • (5,499 rows remain)

III. Data Integrity Audit & Formatting

While the dataset was technically free of nulls, a deep-dive audit was conducted to ensure internal logical consistency between reported funding rounds and actual financial data.

• Financial Reconciliation Audit: We identified a significant mismatch (3,064 records) where the stated funding_rounds did not align with the number of categorized financial columns.

  Click to view the 4-Stage Forensic Audit & Reconciliation

  To achieve 100% logical consistency, the following logic was applied:

  1. Identifying Redundant Aggregates: The venture column was identified as a "noisy aggregate" that double-counted specific rounds (e.g., Series A/B). Removing it reduced logical mismatches to 1,792.

  2. Financial Reconstruction: To preserve data while eliminating noise, we used funding_total_usd as a "truth anchor" and engineered a new feature:

  Unclassified_Funding = Total_Funding - Σ(Categorized_Rounds)

  This isolated grouped or undisclosed investments that were previously hidden within the venture aggregate.

  3. Isolating "Financial Ghosts": Post-reconstruction, the remaining 658 mismatches were identified as records with 0 total funding - essentially missing data disguised as zeros.

  4. Strategic Cleansing & Justification: These 658 records were permanently removed. Rationale: Since venture capital distributions are highly non-normal (skewed), imputing these values would introduce significant bias and distort the relationship between interconnected financial features.

  Final Outcome: Achieved 100% logical consistency across a high-fidelity dataset of 4,841 records.

• Data Sanitation & Normalization: Standardized data types for numerical features and sanitized categorical strings by removing hidden whitespaces and special characters.

• Logical Schema Reordering: Reorganized the dataset structure for improved interpretability: Metadata (Market/Geography) → Global Financial Metrics → Detailed Funding Rounds → Target Status.

IV. Outlier Treatment & Visual Justification

The final stage of data preparation involved addressing extreme values to ensure model robustness. This process was guided by visual diagnostics to distinguish between statistical noise and critical industry signals.

• Temporal Filtering (1990 Cutoff):
  

  
  

  • Observation: The box plot of founded_year revealed a significant "left tail" of legacy companies founded as far back as the early 20th century.

  • Action: Removed records prior to 1990 (approx. 1.7% of the dataset / 84 rows).

  • Rationale: This ensures the analysis remains representative of the modern venture capital and tech landscape, reducing temporal noise that does not reflect current market dynamics.

• Preserving Strategic Extremes (Financial Outliers):
  

  
  
  

  • Observation: Extreme outliers were identified in funding_total_usd and funding_rounds, creating a massive positive skew (Right Tail).

  • Action: These records were intentionally retained.

  • Rationale: In the startup ecosystem, extreme success—such as "Unicorns" or serial fundraises—follows a Power Law distribution. These are not data errors; they are the most critical signals for predicting high-growth outcomes. Removing them would strip the model of its ability to identify the very entities we aim to analyze.

---

5. Descriptive Statistics

Following the data preparation and outlier treatment, this section provides a high-level quantitative and qualitative overview of the finalized dataset (4,841 records).

Numerical Summary (Transposed):

	count	mean	median	std	min	25%	50%	75%	max
founded_year	4757	2005.9	2007	4.09	1990	2004	2007	2009	2014
funding_total_usd	4757	1.83434e+07	5e+06	9.81919e+07	1000	1e+06	5e+06	1.64e+07	5.7e+09
funding_rounds	4757	1.95	1	1.35	1	1	1	2	15
unclassified_funding	4757	4.61351e+06	0	8.27854e+07	0	0	0	1.5074e+06	5.62e+09
seed	4757	193416	0	763443	0	0	0	0	2.5e+07
angel	4757	86483	0	645952	0	0	0	0	3e+07
equity_crowdfunding	4757	1671.22	0	87001.3	0	0	0	0	5.5e+06
convertible_note	4757	10229.8	0	231873	0	0	0	0	1.35201e+07
debt_financing	4757	1.55338e+06	0	2.52692e+07	0	0	0	0	1.2e+09
grant	4757	37656.4	0	1.5033e+06	0	0	0	0	9.98e+07
private_equity	4757	2.16958e+06	0	2.4383e+07	0	0	0	0	7.71e+08
product_crowdfunding	4757	567.58	0	35067.1	0	0	0	0	2.4e+06
round_A	4757	2.11219e+06	0	7.06771e+06	0	0	0	2e+06	2.25e+08
round_B	4757	3.11991e+06	0	9.92575e+06	0	0	0	0	3.5e+08
round_C	4757	2.37028e+06	0	8.63402e+06	0	0	0	0	2e+08
round_D	4757	1.32201e+06	0	8.84972e+06	0	0	0	0	2.5e+08
round_E	4757	514288	0	5.24311e+06	0	0	0	0	2.2e+08
round_F	4757	209741	0	4.98233e+06	0	0	0	0	2.86e+08
round_G	4757	28484.3	0	1.1888e+06	0	0	0	0	6.3e+07
round_H	4757	0	0	0	0	0	0	0	0
status	4757	0.62	1	0.49	0	0	1	1	1

Categorical Overview:

	count	unique	top	freq
market	4757	377	Software	599
country_code	4757	61	USA	3618

Key Statistical Insights:

• Distributional Skewness: The significant gap between the Mean ($18.3M) and the Median ($5M) in total funding confirms a heavy right-skew. This reflects the "Power Law" of the venture capital world, where a few high-value entities significantly influence the overall financial volume.
• Funding Feature Sparsity: While the dataset is rich with 23 features, the individual late-stage columns (Series B and beyond) show high sparsity (75th percentile at 0). This highlights that the majority of startups in the sample operate within early-to-mid funding cycles, making the distinction between "Seed/A" and "Late Stage" a primary factor for investigation.
• Market & Geographic Concentration: With Software (~12.6%) and the USA (~76%) dominating the categorical profile, the dataset is specialized towards the most active hubs of the tech ecosystem, ensuring our model learns from the most mature startup markets.

Feature Interdependence (Correlation Analysis):

To understand how variables influence one another, a Pearson correlation analysis was conducted

Key Correlation Insights:

• Highest Positive Correlation: Among the features analyzed, the number of funding_rounds exhibits the highest relative positive correlation with acquisition status (0.22). While this represents a modest statistical relationship in absolute terms, it stands as the primary signal in our feature set, suggesting that a startup's persistence and ability to secure follow-on rounds are more closely linked to exit potential than other funding metrics.
• Highest Negative Correlation: A negative correlation (-0.24) is observed between founded_year and acquisition status. This reflects a natural "Survival Bias": older companies have had a longer historical window to reach an acquisition event, while newer startups are often still in the early stages of their lifecycle.
• Structural Data Linkage: A high correlation (0.85) exists between funding_total_usd and unclassified_funding. This confirms that unclassified_funding acts as a balancing feature in the dataset's architecture.In addition, this relationship suggests that larger capital raises often involve complex or undisclosed financial structures that are not categorized into standard funding rounds.

---

6. Exploratory Data Analysis (EDA)

This section presents our Exploratory Data Analysis (EDA). Using univariate, bivariate, and multivariate visualizations, we examine the underlying structures, distributions, and key relationships within the data. The following analysis highlights the most significant operational and financial patterns that correlate with a startup's likelihood of being acquired.

To systematically explore these factors, we have structured our analysis around six core questions, each examining the relationship between the target variable and key factors within the dataset.

---

Q1: What is the overall balance between "Acquired" (1) and "Closed" (0) startups in our cleaned dataset?

• Insight: The dataset reveals a distribution of 62% 'Acquired' startups versus 38% 'Closed' startups. This 62/38 split represents a relatively balanced distribution for our target variable.
• Conclusion: This balance is essential for ensuring that our analysis and future predictive models are not heavily biased toward a single dominant outcome.

---

Q2: Does the total investment amount impact a startup's likelihood of being acquired?

• Insight: The log-scale distribution reveals a stark contrast: acquired startups secure a median funding of ~$10M USD, nearly 10 times higher than the ~$1M USD median of closed startups. Furthermore, the 'Acquired' group shows a significantly higher density in the upper funding tiers ($100M+). While Pearson correlation indicates a weak linear relationship (approx 0.074), this metric is artificially dragged down by extreme outliers (heavily funded startups that ultimately failed) and the non-linear nature of venture capital. The visual distribution effectively cuts through this noise, confirming a clear behavioral link between funding volume and status.
• Conclusion: While abundant capital does not guarantee a successful exit, the 10x disparity in medians establishes that crossing a substantial funding threshold is a defining, primary characteristic of startups that achieve acquisition.

---

Q3: Does the number of funding rounds impact the likelihood of being acquired?

• Insight: A comparative analysis of statistical metrics (Mean, Median, and 75th Percentile) reveals a stark contrast in funding momentum. Startups that eventually close face a hard ceiling: 75% of them fail to secure more than two rounds, with a median of exactly 1. Conversely, acquired startups consistently demonstrate stronger momentum, securing at least two rounds (Median = 2). The top 25% of acquired companies scale to 3 or 4+ rounds—a crucial growth stage virtually unreached by failed companies.
• Conclusion: There is a definitive link between consecutive funding rounds and acquisition success. Surviving the initial seed stages to consistently raise multiple rounds (two or more) acts as a rigorous market filter, decisively separating exits from closures.

---

Q4: Which funding rounds are most prevalent among acquired startups?

• Insight: Analyzing the common financial milestones reveals a clear "Exit DNA" heavily reliant on early-to-mid-stage institutional capital. Series A is the most dominant milestone (present in ~40% of acquired startups), closely followed by Series B (~30%). Seed and Series C rounds also serve as core building blocks (~19% each). Interestingly, Debt Financing emerges as a notable secondary instrument (~11%), likely utilized to extend runway and capital flexibility. Conversely, alternative methods like Crowdfunding or Grants are statistical outliers (<1%).
• Conclusion: The typical financial trajectory toward an acquisition is characterized by a progression through standard institutional equity rounds (Seed through Series C). Securing these specific milestones forms the fundamental "Funding Profile" of an acquired startup, whereas very late-stage rounds (Series D and beyond) or alternative funding methods are rarely part of the standard exit pathway.

---

Q5: Which market categories show the highest acquisition rates?

• Insight: An analysis of market categories reveals a clear divergence between sector size and acquisition rates. Broad categories like Software and Mobile, despite dominating in sheer volume (599 and 267 observations, respectively), exhibit lower-tier acquisition rates (~0.669 and ~0.59), signaling highly saturated and competitive environments. Conversely, specialized sectors lead the performance metrics. Analytics boasts the highest acquisition likelihood (~0.842) operating as a focused niche, while Enterprise Software (~0.823) demonstrates remarkable resilience, maintaining a consistently high acquisition rate alongside a substantial volume of activity (215).
• Conclusion: Market size does not positively correlate with a higher probability of acquisition. In fact, startups operating in specialized, B2B-focused niches (such as Enterprise or Analytics) exhibit significantly higher acquisition rates compared to those in broad, mass-market categories. This suggests that delivering targeted solutions offers a more viable path to an exit than operating in high-volume, saturated markets.

---

Q6: How do funding patterns and acquisition outcomes change across different founding years?

• Insight: A chronological analysis reveals three distinct eras of startup evolution across all three key metrics. The early 90s (1990-1995) represent a period of peak acquisition rates and record-breaking median funding, despite a relatively low average number of funding rounds. The subsequent decade (1995-2005) marks a transitional phase: median funding showed a fluctuating upward trend and startups required more consecutive funding rounds to sustain growth, yet overall acquisition rates steadily declined as the market crowded. Finally, the modern era (2005-2013) shows a sharp decrease in both median capital raised and acquisition events, alongside a gradual decrease in average funding rounds.
• Conclusion: The landscape has fundamentally shifted from early, concentrated capital injections to a modern "Lean Startup" approach requiring smaller, incremental funding. Furthermore, the steep drop in recent acquisition rates tangibly visualizes the core principle of startup maturity (Right Censoring or Time-to-Exit bias). As the graph clearly illustrates, younger startups founded closer to 2013 simply have not had an adequate time window to mature, scale, and secure an acquisition compared to their older counterparts.

---

7. Final Conclusion

Our Exploratory Data Analysis (EDA) identifies a distinct pattern predicting a startup’s acquisition probability, driven by the convergence of two primary vectors: Financial Validation and Strategic Positioning.

The analysis reveals that the strongest exit signal is not derived from mere capital accumulation, but from a venture's ability to generate continuous funding momentum. Surpassing the typical "survival ceiling" of the initial two rounds to secure advanced institutional capital (Series A and B) serves as a critical indicator of market trust and scalability.

Crucially, when this financial momentum intersects with specialized, high-value B2B niches - such as Analytics and Enterprise Software - the likelihood of acquisition increases dramatically compared to saturated mass-market sectors.

Ultimately, zooming out to the acquirer's perspective reveals a clear M&A strategy: buyers use robust investment patterns (combining substantial capital volume with continuous funding rounds) as a definitive 'stamp of approval' to validate product quality and de-risk the deal, while specifically targeting niche, B2B startups to instantly buy competitive differentiation in their own markets.

---

8. Limitations

• This analysis is based on observational data and identifies significant statistical associations. However, it does not establish direct causality.
• The dataset covers a specific historical window (predominantly 1990–2013). Since the venture capital ecosystem has evolved significantly since then—with the rise of AI and shifting economic climates—the patterns identified may not fully reflect modern 2026 market dynamics.
• To ensure computational efficiency for this EDA, a sub-sample of the original dataset was utilized. This approach may overlook rare "edge cases" or small niche categories that could offer additional predictive value.
• The dataset lacks critical qualitative indicators—such as founding team experience, burn rate, and net profitability - which are often decisive factors in a company's final exit potential.

---

9. Notebook & Libraries

The full analysis was conducted in a Google Colab environment. The following Python libraries were utilized for data processing, statistical analysis, and visualization:

• Data Manipulation: pandas, numpy
• Visualization: matplotlib.pyplot, seaborn, matplotlib.ticker
• Environment Utilities: google.colab.files

To view the complete data analysis, cleaning process, and visualizations in the official IPYNB file, click the button below:

---

10. Author

Lia Prop

April 2026