Seafood Wholesale Analytics — Data-Driven Inventory Optimization

Customer segmentation, lifetime value analysis, and supply-constrained inventory optimization for a traditional seafood wholesaler.

Overview

This project applies data science techniques to transform a traditional Japanese seafood wholesale business into a data-driven operation. Using historical sales data (2014–2025) and Osaka Central Wholesale Market bonito supply data (2015–2024), I developed analytical models for:

1. Customer Segmentation — B2B/B2C classification + RFM-based sub-clustering
2. Customer Lifetime Value (LTV) Analysis — Identifying "Royal Customers"
3. Inventory Optimization — Counterfactual simulation under supply constraints

Key Result

Under a weekly simulation of supply-constrained scenarios, the proposed priority-based allocation strategy outperforms the conventional First-Come-First-Served (FCFS) approach:

Supply Ratio (α)	FCFS Profit (JPY)	Proposed Profit (JPY)	Improvement
0.5 (severe)	20,772,650	24,156,917	+16.29%
0.6	24,821,450	28,543,804	+15.00%
0.7	28,856,817	32,641,803	+13.12% (+¥3.78M)
0.8	32,924,960	36,562,615	+11.05%
0.9	36,994,360	40,410,971	+9.24%

At α=0.7 (severe shortage), the proposed model achieves +13.12% gross profit improvement (approx. ¥3.78 million annually) by prioritizing high-LTV "Royal Customers" during supply shortages.

Project Structure

├── data/
│   ├── sample_sales_data.csv          # Anonymized sales transactions
│   └── market_supply_data.csv         # Osaka bonito market supply (2015-2024)
├── notebooks/
│   ├── 01_data_exploration.py         # Exploratory Data Analysis
│   ├── 02_customer_clustering.py      # B2B/B2C segmentation + sub-clustering
│   ├── 03_ltv_analysis.py             # Customer Lifetime Value analysis
│   └── 04_inventory_optimization.py   # Supply-constrained allocation simulation
├── results/
│   ├── figures/                       # Auto-generated visualizations
│   └── simulation_results.csv         # Simulation output
├── requirements.txt
└── README.md

Dataset

Sales Data

Anonymized transaction records from a seafood wholesale company in Osaka, Japan.

Column	Description
date	Transaction date (2014–2025)
customer_code	Anonymized customer identifier
customer_name	Anonymized customer name
product_code	Anonymized product identifier
product_name	Anonymized product name
quantity	Units sold
unit_price	Selling price per unit (JPY)
cost_price	Cost price per unit (JPY)
sales_amount	Total transaction value (JPY)
gross_profit	Gross profit (JPY)
product_type	Product type code
product_category	Product category code

Market Supply Data

Monthly frozen bonito (skipjack tuna) arrivals at Osaka Central Wholesale Market.

Column	Description
market_volume_kg	Monthly arrival volume (kg)
market_value_yen	Monthly total value (JPY)
market_avg_price	Average price per kg (JPY)

Analysis

1. Exploratory Data Analysis (01_data_exploration.py)

• 79,935 valid transactions across 303 customers and 1,213 products
• Strong Pareto effect: Top 10 customers account for 67.6% of total revenue
• Clear seasonality aligned with bonito fishing seasons
• Sales volume loosely correlates with market supply

2. Customer Clustering (02_customer_clustering.py)

• B2B/B2C Classification: Feature-based Logistic Regression using order patterns, volume, and category diversity
• B2B Sub-clustering: 3 segments via K-Means (RFM features)
• B2C Sub-clustering: 4 segments via K-Means (RFM features)
• Optimal K determined via Elbow Method + Silhouette Score

3. LTV Analysis (03_ltv_analysis.py)

• Gross-profit-based LTV calculation across full customer history
• Strong Pareto distribution: Top 20% of customers = 94.4% of total LTV
• Royal Customers: 56 customers with mean LTV of ¥19.4M, avg tenure 7.9 years
• Predictive model: Random Forest achieves R² = 0.83 for LTV prediction from early behavior

4. Inventory Optimization (04_inventory_optimization.py)

• Weekly granularity simulation for precise supply shock capture
• Counterfactual design: 2015–2023 training, 2024 validation
• Two strategies compared:
  • FCFS: First-Come, First-Served (conventional)
  • Proposed: LTV-based priority allocation to Royal Customers
• Result: +13.12% profit improvement at α=0.7, functioning as "intertemporal arbitrage"

Getting Started

# Clone the repository
git clone https://github.com/YOUR_USERNAME/seafood-wholesale-analytics.git
cd seafood-wholesale-analytics

# Install dependencies
pip install -r requirements.txt

# Run the analysis notebooks (from the notebooks/ directory)
cd notebooks
python 01_data_exploration.py
python 02_customer_clustering.py
python 03_ltv_analysis.py
python 04_inventory_optimization.py

Note: Scripts use # %% cell markers compatible with VS Code / Jupyter interactive mode. To convert to .ipynb:

pip install jupytext
jupytext --to notebook notebooks/01_data_exploration.py

Tech Stack

• Language: Python 3.10+
• Data Processing: pandas, NumPy
• Machine Learning: scikit-learn (K-Means, Logistic Regression, Random Forest, Gradient Boosting)
• Visualization: matplotlib, seaborn
• Simulation: Custom counterfactual allocation engine

Context

This project was developed during a Data Science Internship at a seafood wholesale company in Osaka, Japan (July 2025 – Present). The company had no prior data analytics infrastructure, and this project served as the foundation for their digital transformation initiative.

License

This project is licensed under the MIT License — see LICENSE for details.

Disclaimer: All data in this repository is anonymized. No actual business data is included.