Skip to content

README.md

Latest commit

 

History

History
203 lines (153 loc) · 8.41 KB

README.md

File metadata and controls

203 lines (153 loc) · 8.41 KB

VSP Vision — Frame Assortment Intelligence

Demand forecasting + product management analysis for Visionworks eyewear retail
XGBoost · Python · 162 SKUs · 3 Brands · 2 Regions · Sep 2023–Aug 2024

View the full HTML analysis report here:

👉 Open the Full VSP Frame Assortment Report

The report includes:

  • Full exploratory analysis
  • Model methodology
  • Feature importance visualizations
  • Forecast outputs
  • All 14 prioritized action items
  • Interactive charts and tables

Overview

An end-to-end demand forecasting pipeline built for Visionworks frame assortment planning. Starting from raw monthly sales data, this project enriches product attributes, engineers 33 temporal features, trains and compares three ML models, and translates model output into 14 concrete, data-backed product management recommendations.

The goal: replace intuition-driven assortment decisions with a repeatable, leakage-free framework that predicts SKU-level monthly demand and surfaces where the assortment has structural gaps.

Results

Metric Value
Model XGBoost (Tuned)
R² — Log Scale (primary) 0.737
R² — Raw Units 0.542
Cross-Validated R² 0.738
MAE 26.4 units / month
Features 33
Training period Sep 2023 – May 2024
Test period Jun – Aug 2024
Total units analyzed 910,853

Project Structure

vsp-frame-assortment/
│
├── data/
│   ├── detailed_attributes.csv        # Style-level product attributes
│   └── final_demand.csv               # SKU × Region × 12 monthly sales columns
│
├── docs/
│   ├── Final Report VSP.pdf           # Final written report
│   ├── VSP Presentation.pdf           # Slide deck presentation
│   └── methodology.md                 # Methodology write-up
│
├── notebooks/
│   ├── Demand_and_Cluster_Prediction.ipynb   # Demand forecasting + clustering
│   ├── Descriptive-analytics.ipynb           # Exploratory data analysis
│   ├── Enriched_demand.ipynb                 # Data enrichment + feature prep
│   └── VSP_XGboost_finetuned.ipynb           # Final tuned XGBoost model
│
├── outputs/
│   ├── VSP_Report_Updated.html        # Updated interactive HTML report
│   ├── vsp_report.html                # Original HTML report
│   └── action_items.docx              # Prioritized PM action plan
│
├── index.html                         # Portfolio landing page
├── requirements.txt
├── .gitignore
└── README.md

The Business Problems

Problem 1 — The $150–$199 tier is over-assorted.
64 SKUs (39% of the assortment) concentrated in a single price band generating only $4,017/SKU annually — 37% below the $200–$249 sweet spot at $6,326/SKU.

Problem 2 — Female patients are structurally underserved.
Female-labeled SKUs account for just 15.8% of sales despite having identical per-SKU productivity to male SKUs ($4,783 vs $4,817). The gap is supply, not demand.

Methodology

Data Preparation

  • Two source files merged on StyleCode: demand file (1,240 SKU–region rows × 12 monthly columns) + attributes file (220 styles)
  • Cleaned 3,055 null values: size typos (3.0 → 53mm), casing mismatches, missing attributes
  • Reshaped to long format: 14,880 rows (one per SKU–Region–Month)
  • Log-transformed the sales target log(1 + Sales) — raw skewness = 2.56
  • Time-based split: train Sep 2023–May 2024 / test Jun–Aug 2024

Feature Engineering — 33 Features, All Leakage-Free

Category Features
Lag Sales 1, 2, 3, 6 months prior
Rolling averages 3-month and 6-month rolling mean
Expanding statistics SKU mean, max, cumulative (since launch)
Trend MoM growth (clipped ±200%), volatility, SKU vs brand peers
Time Month, quarter, year, Jan–Mar peak-season flag
Price Log price, price tier, margin ratio, price vs brand average
Attributes Brand, Region, Gender, Shape, Material, SunOptical (label encoded)

Key decision: StyleCode was dropped. Keeping it inflated R² from 0.52 → 0.69 by memorizing SKU identities rather than learning demand patterns. A model that knows which SKU it's predicting cannot generalize to new styles.

Model Comparison

Model R² (Log) R² (Raw) MAE
Linear Regression 0.310 −18.0 55.6
Random Forest 0.520 0.514 26.9
XGBoost (baseline) 0.580 0.523 26.9
XGBoost (tuned) 0.737 0.542 26.4

Tuned via RandomizedSearchCV across 40 combinations, 3-fold CV.
Best params: n_estimators=800, max_depth=4, learning_rate=0.01, subsample=0.7, reg_lambda=3

Key Findings

What drives demand: Sales trajectory beats all static product attributes. Top 5 features by importance: Rolling 6-Month Avg Sales → SKU Max Sales → SKU Avg Sales → Brand → Optical/Sun. Shape, material, and size have weak individual signal but strong interaction effects.

Assortment gaps:

  • Oval frames — 1 SKU, 9,432 units, highest avg/SKU of any shape. Proven demand, no supply.
  • Titanium — 5 SKUs at $9,852/SKU (2× Acetate). NIKE 8045 and 8154 are both top-10 SKUs.
  • Cat Eye — 2 SKUs total. Nationally growing trend. Visionworks has no real presence.
  • CK Sun — 80% dead or slow SKUs, $1,197/SKU. Worst performing brand in portfolio.
  • Lacoste in AMER — 37.6% of EMEA but only 14.6% of AMER. Their $200–$249 tier = $10,005/SKU.

Seasonality: Sales jump 73–80% from December to March (insurance plan resets). New styles must be floor-ready by December 15 or miss ~40% of their first-season window.

Top 10 SKUs

Rank Style Brand Region Shape Material Price Units
1 L2707 Lacoste Optical EMEA Square Acetate $199.95 20,417
2 L2876 Lacoste Optical EMEA Square Acetate $220.00 14,317
3 NIKE 7125 Nike Optical AMER Rectangle Bio Inj-G820 $249.00 14,219
4 NIKE 7246 Nike Optical AMER Rectangle Acetate $259.00 14,096
5 L2707N Lacoste Optical EMEA Square Acetate $199.95 13,443
6 NIKE 8045 Nike Optical AMER Rectangle Titanium $295.00 13,048
7 L2271 Lacoste Optical EMEA Rectangle Metal $225.00 12,086
8 NIKE 8154 Nike Optical AMER Rectangle Titanium $295.00 11,767
9 NIKE 7090 Nike Optical AMER Square Bio Inj-G820 $285.00 11,583
10 NIKE 5538 Nike Optical AMER Square Bio Inj-G820 $135.00 11,129

Priority Action Items

# Priority Action Justification
1 🔴 HIGH Cut 10–15 SKUs from $150–$199 tier $4,017/SKU vs $6,326/SKU — 37% gap
2 🔴 HIGH Add 4–6 Lacoste styles to AMER ($200–$249) $10,005/SKU — best brand×tier combo
3 🔴 HIGH Floor Hero SKUs by Dec 15 73–80% sales jump Jan–Mar
4 🔴 HIGH Expand Titanium to 8–10 SKUs $9,852/SKU, nearly 2× Acetate
5 🟡 MED Cut CK Sun from 15 → 5–6 SKUs $1,197/SKU, 80% dead
6 🟡 MED Reduce Lacoste Suns; EMEA-only 17/18 SKUs dead, 84% EMEA
7 🟡 MED Add female-specific Nike + CK styles Equal productivity, 30 vs 69 SKUs
8 🟡 MED Build dedicated EMEA buy list Current top-10 is 9/10 Nike AMER
9 🟡 MED Separate AMER/EMEA size grids AMER: 54mm peak, EMEA: 53mm peak
10 🟡 MED Rationalize CK Optical to 20–25 SKUs $3,809/SKU — lowest optical brand
11 🔵 LOW Test 2–3 Oval optical styles 9,432 avg/SKU — highest of any shape
12 🔵 LOW Test 2–3 Cat Eye styles 2 SKUs only, growing national trend
13 🔵 LOW Extend model to 24+ months of history +0.04–0.08 R² estimated gain
14 🔵 LOW Implement Dec 15 launch calendar Late arrivals miss 40% of seasonal peak

Setup

git clone https://github.com/yourusername/vsp-frame-assortment.git
cd vsp-frame-assortment
pip install -r requirements.txt

Open notebooks in order:

notebooks/01_data_cleaning.ipynb
notebooks/02_eda.ipynb
notebooks/03_xgboost_model.ipynb

Tech Stack

Python 3.11 pandas numpy XGBoost 2.0 scikit-learn matplotlib Jupyter

Graduate-level data analytics project — VSP Vision / Visionworks frame assortment analysis.