February 2024 - September 2024
This repository documents my work as an AI Developer at AVILEN on the 日本ハム project, where I focused on data analysis and predictive modeling for market trends and pricing.
Avilen-portfolio
├── 01_data_cleaning # Data preprocessing pipelines
├── 02_cross_market_export_analysis # Export market visualization
├── 03_販売単価分析 # Sales price analysis
├── 04_販売単価_vs_○○○_analysis # Price correlation studies
├── 05_predicting_large_changes_in_外貨 # Currency change prediction
├── 06_predicting_出回り数量_LightGBM # Distribution volume forecasting
├── 07_predicting_フード販売単価 # Food price prediction
└── 08_research_task # Industry AI implementation research
Tools: Python, pandas
Processed three types of data sources:
-
Container Freight Charges
- Converted PDF format into structured CSV data
- Standardized date formats to yyyymm
- Created consistent numeric columns
-
Market Conditions Data
- Consolidated data from Brazil, Thailand, USA, and China
- Standardized currency representations
- Unified missing value handling
-
Brazil Export Analysis
- Transformed export data into analyzable time series
- Processed nested year data structures
- Implemented country name translations
Tools: Python, matplotlib
Created visualizations comparing:
- US exports (shifted 1 month forward)
- Brazilian exports (scaled and shifted 1 month forward)
- US inventory levels
- Foreign currency rates
Generated analysis for key markets:
- Philippines
- China
- Angola
- Iraq
- South Africa
Integrated six economic indicators:
- Corporate Price Index (企業物価指数)
- Wage Data (賃金データ)
- Consumer Price Index (消費者物価指数)
- Interest Rates (金利)
- GDP Data
- GDP Growth Rate (GDP成長率)
Analyzed relationships between:
- Sales unit prices and consumer prices
- Shipping volumes and inventory metrics
- Market indicators with lag effects
Key findings from detrended analysis:
- Earlier years (2012-2017): Predominantly negative detrended values
- Recent years (2022-2024): Higher volatility with both positive and negative deviations
Analyzed different thresholds for predicting significant currency price changes:
| Threshold | Precision | Recall | F1 Score | Accuracy |
|---|---|---|---|---|
| 0.5 | 0.6667 | 0.3784 | 0.4828 | 0.7000 |
| 1.0 | 0.7778 | 0.1892 | 0.3043 | 0.6800 |
| 1.5 | 0.6667 | 0.1081 | 0.1860 | 0.6500 |
| 2.0 | 1.0000 | 0.0541 | 0.1026 | 0.6500 |
| 2.5 | 1.0000 | 0.0270 | 0.0526 | 0.6400 |
| 3.0 | 1.0000 | 0.0270 | 0.0526 | 0.6400 |
Tools: LightGBM
Implemented various feature combinations with results:
- Basic Inventory Lags: MAPE 2.15%
- Rolling Statistics: MAPE 2.13%
- Inventory Differences: MAPE 2.13%
- Currency Lags: MAPE 2.14%
- Currency Differences: MAPE 2.11%
- Currency-Inventory Interactions: MAPE 2.40%
- Categorical Features: MAPE 2.13%
- Derived Inventory Metrics: MAPE 2.15%
- Conditional Features: MAPE 2.06%
- Final Combined Model: MAPE 2.04%
Developed LightGBM model with feature engineering:
- Lag features for key metrics (2-6 months)
- Rolling statistics
- Interaction features between currency rates and inventory levels
Final model achieved MAPE of 2.04% on test data starting from January 2021.
Researched AI implementation in major food companies:
Tyson Foods Implementation:
- AWS machine learning for chicken tray detection
- Quantified Result: 15,000 hours of skilled labor saved annually in one facility
- Automation Program: $1.3 billion investment over three years
- Results: 48% EPS growth to $2.87 in Q1 2022, 24% sales increase
JBS Implementation:
- AI-driven carcass sorting system
- AI cameras for tender steak processing quality assessment in Canada
Sysco Implementation:
- Partnership with iFoodDS for traceability solutions
- Focus on FSMA 204 compliance
Project feedback from Kota Kobayashi, Data Scientist at AVILEN:
"時系列データを用いたコード実装のアウトプットはいつも期待値を超えていました。抽象的な業務も自身でタスクに分解して期待以上のアウトプットを出せます。"