How I Built a Python Demand Forecasting Model That Saved $18,000

#datascience #machinelearning #python

Emergency procurement was happening 23 times per month at the automotive parts company where I work. The purchasing team was reactive rather than proactive, causing stockouts and premium shipping costs.

I built a demand forecasting model that reduced emergency procurement by 65% and saved $18,000 annually.

The Problem

The team was restocking after stock hit zero rather than before. By the time they noticed, supplier lead times meant there was already a gap. Emergency orders cost more and disrupted operations.

The Solution

Data Collection

I pulled 13 months of stock movement data from our SQL inventory database — roughly 195,000 rows across 512 SKUs.

Feature Engineering

I engineered features including:

Rolling average consumption (7-day and 30-day)
Days of stock remaining
Stock buffer ratio
Seasonal indicators (month, day of week)

Model Selection

I tested three approaches:

Simple threshold rule (too many false positives)
Gradient Boosting (good but lower precision)
Random Forest (best balance)

Random Forest gave 78% precision and 76% recall.

Deployment

The model runs daily at 7 AM and outputs a prioritized reorder list directly to the operations team as an Excel report.

Results

Metric	Before	After
Emergency orders/month	23	8
Reduction	-	65%
Annual savings	-	$18,000

Key Takeaway

The model's accuracy mattered less than the cross-functional review process. Consistent weekly meetings between procurement and data teams drove the real behavior change.
(https://danielnnadi.github.io/DanielNnadi/article-1-sql-optimization.html))