rotemvahava/flipfinder-dataset

🏠 FlipFinder USA

Identifying Real Estate Investment Opportunities Across the United States

Authors: Dan & Rotem | HuggingFace: @rotemvahava

---

📋 Project Overview

This project transforms a synthetically generated real estate dataset into a focused investment screening tool. Using Exploratory Data Analysis (EDA), we engineered a multi-class target variable (investment_label) that classifies each property into a specific investment strategy — Flip, BRRRR, Buy and Hold, Bad Investment, or Uncategorized — based on how it is priced relative to its immediate local market, its rental yield, and its neighborhood quality.

This dataset powers the FlipFinder AI recommendation app.

---

❓ The Question We Want to Answer

Is this property a good investment — and if so, what type: Flip, BRRRR, or Buy-and-Hold?

---

📦 Dataset

• Source: synthetically generated with Qwen/Qwen2.5-3B-Instruct (HuggingFace) using 5 prompting techniques
• Raw size: ~11,400 rows × 22 features
• Cleaned size: 10,738 rows × 43 features (22 original + engineered metrics, scores, and labels)
• Coverage: 7 US states · 22 cities · 177 ZIP codes · 4 property types
• Type: Numeric + categorical tabular data

Key Features

Feature	Type	Description
state	Categorical	US state (full name, e.g. Texas, Ohio, Florida)
city	Categorical	City where the property is located
zipcode	Categorical	5-digit US ZIP code
listed_price	Numerical	Listed asking price in USD
market_estimate	Numerical	Model-estimated market value
rent_estimate	Numerical	Estimated monthly rent
sqft	Numerical	Interior living area in square feet
price_per_sqft	Numerical	Price per square foot (listed_price / sqft)
lot_area_acres	Numerical	Lot size in acres
bedrooms	Numerical	Number of bedrooms
bathrooms	Numerical	Number of bathrooms (incl. half baths)
year_built	Numerical	Year of construction
days_on_market	Numerical	Days the property has been listed
property_type	Categorical	Single Family, Multi-Family, Condo, Townhouse
niche_overall_grade	Categorical	Overall neighborhood grade (A+ to B-)
school_rating	Categorical	Local school quality grade
crime_safety_rating	Categorical	Neighborhood safety grade
housing_rating	Categorical	Housing market quality grade
nightlife_rating	Categorical	Local amenities / nightlife grade
latitude / longitude	Numerical	Geographic coordinates
investment_label	Categorical	Multi-class target: Flip / BRRRR / Buy and Hold / Bad Investment / Uncategorized

---

🎯 Target Variable — investment_label

The dataset has no built-in classification target, so we engineered our own using a comparables-based valuation rather than relying on the synthetic market_estimate field.

For each property:

• Calculate the local median price per square foot
• Group by ZIP + Bedroom + property_type to find local comps (minimum 5 properties)
• Fall back to ZIP + Bedroom, then to the 3-digit ZIP prefix + Bedroom if a group is too small
• Multiply the local median PPSq by the property's sqft to get its ARV (After-Repair Value)
• Derive five investment metrics: price_vs_market, gross_yield, cap_rate, price_to_rent_ratio, property_age

Labels are then assigned by a priority-ordered rule system:

Label	Rule (summary)
🔴 Bad Investment	overpriced vs comps, or yield < 4%, or > 120 days listed, or neighborhood below B-
🟡 Flip	≥ 15% below ARV, not new build (< 2020), ≤ 120 days listed, not a condo
🟠 BRRRR	≥ 10% below ARV, gross yield ≥ 8%, price-to-rent < 15, B-grade or better, not a condo
🟢 Buy and Hold	yield 5-8%, 20-120 days listed, built after 1990, strong neighborhood/school/safety grades
⚪ Uncategorized	meets none of the above

Distribution: Bad Investment 5,347 (49.8%) · Uncategorized 2,812 · Buy and Hold 1,568 · Flip 675 · BRRRR 336.

---

🧹 Section 2: Data Wrangling

The generated data arrived complete — zero missing values, zero fully duplicated rows — so cleaning focused on verifying integrity and enforcing consistency rather than imputation:

• Numeric validity: all prices, areas, rents, and days-on-market positive; bedrooms/bathrooms 1-12; build years 1800-2024; valid 5-digit ZIPs; all coordinates inside the continental US.
• price_per_sqft fix: 236 rows (~2%) where price_per_sqft disagreed with listed_price / sqft were recomputed directly.
• state: collapsed a stray "TX" into "Texas" (8 → 7 canonical state names).
• property_type: folded two non-standard labels into the 4 canonical categories (6 → 4 types).
• Grades & bathrooms: all neighborhood grades within the 6 allowed letters; all bathroom counts on the .0/.5 convention.
• Coordinate consistency: every property's coordinates fall within its own state's real bounds.
• Outliers: extreme values removed via box-plot inspection and logical filters.

---

🔄 Section 3: Data Transformation

Both listed_price and price_per_sqft were strongly right-skewed. We applied a log transformation with np.log1p, which reduced skewness from ~1.70 to near 0, normalizing the distributions for downstream modeling. The five ordinal letter-grade columns were also converted to GPA-style numeric scores. All EDA continues to use the original dollar values for interpretability.

---

⚙️ Section 4: Feature Engineering

The core of the project. From the comps-based ARV we engineered the five investment metrics above, then applied the priority-ordered rule system to assign each property its investment_label. This multi-class approach reflects how a real investor evaluates a property — not just whether it is cheap, but which specific strategy it best supports given its price, yield, condition, and neighborhood context.

---

📊 Section 5: Descriptive Statistics

A correlation review across the numeric features and the target confirmed that no single feature correlates strongly with investment_label — the investment signal is driven by a combination of variables (below-comps pricing + adequate yield + neighborhood quality), not any one feature alone. This is exactly why a comps-based, rule-driven approach outperforms any single-feature filter.

---

📈 Section 6: Univariate & Bivariate Analysis

• Price distribution is strongly right-skewed, with most properties clustered in the mid-market range — the sweet spot for realistic investment candidates.
• PPSq deviation by label validates the target: Flip properties cluster to the left of the -15% line and BRRRR between -10% and -15%, confirming the thresholds behave as designed.
• Gross yield by label shows BRRRR sitting highest (≥ 8%), Buy and Hold in the 5-8% band, and Bad Investment below the 4% cutoff — a clean separation matching the labeling rules.
• Investment mix by state reveals geographic variation in opportunity density despite the comps-based (relative) labeling.

---

❓ Section 7: Research Questions

• Q1 — Which ZIP codes have the most Flip opportunities? Ranked ZIPs with ≥ 20 listings by Flip rate; a handful of ZIPs concentrate opportunities well above the dataset average.
• Q2 — Which states offer the most opportunities? Ranked all 7 states by combined Flip + BRRRR rate. Ohio leads at 12.7%, with two ZIP codes in the national top 5.
• Q3 — What does a typical opportunity look like? Flip and BRRRR candidates cluster in the $150K-$400K, mid-size Single/Multi-Family segment.
• Q4 — Does neighborhood quality relate to investment type? A+/A areas favor Buy and Hold (43% / 35%); B and B- areas produce the most Flips (~10-12%).
• Q5 — What price range and property type produce the most opportunities? The affordable, Single/Multi-Family segment; condos are structurally excluded from active strategies.

---

📝 Section 8: Key Findings

Finding 1 — The investment signal is relative, not absolute. No single raw feature predicts the label well alone. The signal emerges from a property being priced below its local comps, having adequate rental yield, and meeting neighborhood-quality thresholds — which is why the comps-based approach beats any single-feature filter.

Finding 2 — Price range and property type define the opportunity profile. Flip and BRRRR opportunities concentrate between $150K and $400K in Single Family and Multi-Family homes. Condos are structurally excluded from active strategies. Investors should focus their search in this price-type combination for the highest opportunity density.

Finding 3 — Neighborhood grade predicts strategy type, not opportunity existence. A+ and A neighborhoods favor Buy and Hold (43% and 35%), while B and B- neighborhoods produce the most Flip opportunities (~10-12%). Bad Investment appears at every grade level (44-55%), confirming that overpricing is not limited to any specific neighborhood tier.

Finding 4 — Ohio is the standout cash-flow market. Ohio leads all states at a 12.7% combined Flip + BRRRR rate, with two ZIP codes in the top 5 nationally. For BRRRR-focused investors, Ohio ZIP codes are the primary target.

Finding 5 — The Bad Investment rate reflects conservative labeling. Nearly half the dataset (49.8%) was labeled Bad Investment, primarily driven by the tight price clustering of the synthetic data. In real market data with more price variance, Flip and BRRRR rates would likely be higher.

---

⚠️ Limitations

• Synthetic data — findings demonstrate methodology rather than real market truth
• Tight price clustering — the generator favors round-number prices in narrow ranges, reducing variance
• Comps limited to the dataset — ARV is computed from these 10,738 properties, not a full MLS database
• No transaction costs — renovation, closing, and carrying costs are not reflected in the labels
• Static snapshot — no temporal dimension; market conditions change over time
• Grade consistency — the generator tended to assign uniform grades across the five neighborhood dimensions, reducing their independent signal

---

📁 Repository Contents

File	Description
flipfinder_final.csv	Cleaned, transformed, labeled dataset (10,738 × 43)
Dan & Rotem Part 1 - Synthetic Data Generation.ipynb	How the raw data was generated
Dan & Rotem Part 2 - Dataset & EDA.ipynb	Full EDA notebook with all code and explanations
README.md

README.md

🚀 How to Run

1. Open the .ipynb files in Google Colab or Jupyter Notebook
2. Run all cells from top to bottom
3. All cleaning, transformation, feature engineering, and visualizations are generated automatically

---

Project by Dan & Rotem | FlipFinder USA | 2026