Predictive Analytics Dashboard

The Challenge

ShopSense was flying blind. With over 12,000 SKUs across five warehouses, their inventory decisions were based on spreadsheets, gut feeling, and last year's sales data. The result: chronic overstocking of slow movers and stockouts on their best sellers.

Overstock was costing them $2.1M annually in carrying costs. Stockouts were worse — an estimated $3.4M in lost revenue from customers who bought elsewhere when items were unavailable. Their planning team knew they needed forecasting, but previous attempts with off-the-shelf tools had plateaued at 68% accuracy — not reliable enough to change how they operated.

They came to us asking for a dashboard. What they actually needed was a forecasting engine that could account for seasonality, promotions, and cross-product effects — then a dashboard to make it actionable.

Our Approach

We spent the first three weeks deep in ShopSense's data: two years of transaction history, promotional calendars, supplier lead times, and warehouse capacity constraints. We found that their demand patterns were heavily influenced by factors the off-the-shelf tools couldn't model — promotional cannibalization between similar products, regional weather effects on seasonal items, and supplier reliability patterns.

We built a custom time-series forecasting stack using Prophet as the base model, enhanced with gradient-boosted residual models that capture the cross-product and promotional effects. The ensemble approach gave us the accuracy ceiling we needed.

The dashboard was designed around decisions, not just data. Every view answers a specific question: 'What should I reorder this week?', 'Which warehouses are at risk of stockout?', 'How will this promotion affect inventory?'

The Solution

The production system runs on AWS: a data pipeline built with Apache Airflow pulls daily sales data, enriches it with external signals (weather, holidays, competitor pricing where available), and feeds the forecasting models.

The Next.js dashboard delivers real-time inventory visibility with drill-down from company-wide to individual SKU level. Automated alerts fire when forecast confidence drops below thresholds or when reorder points are approaching. The system generates weekly purchase recommendations that account for supplier lead times and warehouse capacity.

We built a simulation mode that lets the planning team test 'what-if' scenarios — 'What happens to inventory if we run a 30% off sale next month?' — before committing to decisions.

Results

Lessons Learned

Accuracy isn't everything — we hit 94% in testing but deliberately backed off to 92% in production because the last 2% required features that updated too slowly (quarterly supplier data). A model that's 92% accurate with real-time inputs beats a 94% model that's stale half the time. Choosing the right accuracy-freshness tradeoff was key.

What We'd Do Differently

With what we know now, we'd invest in the simulation mode earlier. The planning team didn't start trusting the system until they could test their own 'what-if' scenarios and see the model's reasoning. We built it in month three — it should have been in the MVP.