AI-Powered Demand Forecasting That Helped a Seasonal Retailer Stop Over-Ordering and Start Selling Through

A seasonal outdoor recreation retailer with three store locations and twenty-five employees had spent fifteen years building a strong regional reputation for kayaks, camping gear, ski equipment, and outdoor apparel. The business thrived during peak seasons but carried a chronic overstock problem that consumed warehouse space, tied up cash, and forced deep markdowns every transition period. The owner estimated that end-of-season clearance events erased twenty to thirty percent of the margin earned during peak selling weeks.

The retailer's purchasing process relied entirely on the owner's experience and vendor recommendations. Orders were placed based on what sold well the previous year, adjusted by gut feel for trends. There was no systematic analysis of sell-through rates, no modeling of seasonal curves, and no mechanism to adjust purchasing mid-season based on actual demand signals. The result was predictable: some categories sold out too early while others sat on shelves until they were marked down to move.

JSG was engaged to design and deploy an AI-powered demand forecasting system that would bring data-driven precision to the retailer's purchasing decisions, reduce overstock waste, improve sell-through rates, and minimize margin-destroying markdowns.

Client: Seasonal outdoor recreation retailer operating 3 store locations in a mountain/lake region. Fifteen years in operation with a loyal customer base and strong seasonal revenue swings concentrated in summer (May-September) and holiday (November-December) periods.

Employee Size: 25 employees

Industry: Retail & E-Commerce

Services: - Machine Learning Demand Forecasting - Seasonal Pattern Modeling & Analysis - Automated Purchasing Recommendation Engine - Markdown Optimization & Sell-Through Analytics

The retailer's overstock problem had compounded over fifteen years of experience-based purchasing. The owner was a deeply knowledgeable outdoors enthusiast who understood his products and his customers, but his purchasing instincts were calibrated to a smaller, simpler version of the business. With three locations, hundreds of product lines, and seasonal swings that varied by category, location, and weather pattern, human intuition alone could not optimize inventory across the full matrix of variables.

First, purchasing decisions were made once per season with little ability to course-correct. The owner placed the bulk of each season's inventory orders four to five months in advance based on the prior year's sales and vendor early-order discounts. Once product arrived, the only lever available was pricing -- if kayaks were not moving, the answer was a markdown. There was no system to detect slow sell-through early enough to shift marketing spend, adjust pricing surgically, or redirect inventory between locations.

Second, seasonal patterns varied significantly by product category and location, but the purchasing approach treated the business as a monolith. Ski equipment at the mountain-adjacent store followed a sharply different demand curve than camping gear at the lakeside location. The owner acknowledged that he tended to over-order categories he was personally enthusiastic about (fly fishing, backcountry camping) and under-order categories he found less interesting (casual outdoor apparel) regardless of actual demand data.

Third, the retailer had three years of point-of-sale transaction data sitting unused in disconnected systems. Each store's POS captured detailed transaction records -- product, quantity, price, date, time -- but no one had ever analyzed this data in aggregate. The information needed to build accurate demand models already existed; it simply had never been extracted, cleaned, or modeled.

Fourth, markdown decisions were reactive and blunt. When the owner decided it was time to clear seasonal inventory, he applied blanket percentage discounts across entire categories rather than targeting specific SKUs based on their individual sell-through trajectory. Products that might have sold at full price with two more weeks of shelf time were swept up in the same clearance event as genuine dead stock, unnecessarily destroying margin.

The retailer needed:

• A demand forecasting model trained on actual historical sales data across all 3 locations
• Seasonal pattern detection that captured category-level and location-level variation
• Automated purchasing recommendations with quantity guidance by SKU and location
• Mid-season sell-through tracking with early warning alerts for slow-moving inventory
• Markdown optimization that targeted discounts at the SKU level based on individual performance

JSG designed and deployed an AI-powered demand forecasting system that transformed the retailer's purchasing process from gut-driven seasonal bets into a data-informed, continuously adjusting inventory strategy.

Key Components

Historical Data Consolidation & Cleaning JSG extracted three years of transaction data from all three POS systems -- over 180,000 individual transactions -- and consolidated them into a unified dataset. Data cleaning addressed inconsistencies in SKU naming, handled returns and exchanges, normalized promotional pricing effects, and tagged each transaction with location, category, sub-category, and seasonal period metadata.

AI Demand Forecasting Model Azure OpenAI was trained on the consolidated dataset to build demand forecasting models at the SKU-location level. The model incorporated seasonal decomposition (identifying recurring summer peaks, holiday surges, and shoulder-season patterns), year-over-year growth trends, category lifecycle curves, and promotional lift effects. Forecasts were generated at weekly granularity for the upcoming twelve months, with confidence intervals that widened appropriately for longer-horizon predictions.

Automated Purchasing Recommendations N8N workflows translated the demand forecasts into actionable purchasing recommendations. Before each buying season, the system generated a recommended purchase order by SKU and location, including suggested quantities, optimal order timing to align with vendor lead times, and estimated sell-through probability at full price. The owner reviewed and adjusted recommendations before placing orders, but for the first time had a data-driven starting point rather than a blank spreadsheet.

Mid-Season Sell-Through Tracking Power BI dashboards tracked actual sell-through rates against forecasted rates in real time, updated daily from POS data feeds. The system flagged SKUs that were selling significantly faster or slower than predicted, triggering automated alerts through N8N. Fast sellers triggered reorder recommendations (when vendor lead times allowed), while slow movers triggered targeted markdown suggestions before the end-of-season clearance window.

Markdown Optimization Engine Azure OpenAI powered a markdown optimization model that recommended discount levels for individual SKUs based on remaining inventory, weeks until season end, historical price elasticity, and storage cost. Instead of blanket 30%-off clearance events, the system recommended graduated markdowns -- 10% for items that needed a slight push, 25% for genuine slow movers, and full-price continuation for items with healthy sell-through trajectories that simply needed more time.

## Quantifiable Impact

• End-of-season overstock reduced by 40%, from an average of $215,000 in clearance inventory to $129,000
• Full-price sell-through rate improved from 58% to 73% across all categories
• Markdown losses reduced by 35%, saving an estimated $67,000 annually in margin preservation
• Purchasing accuracy improved -- stockout incidents during peak season decreased by 52%
• Reorder recommendations during mid-season captured an estimated $41,000 in revenue from fast-selling items that previously would have stocked out
• Inventory carrying costs reduced by 22% due to tighter purchasing alignment with actual demand

• Cloud: Microsoft Azure
• AI & Automation: Azure OpenAI (demand forecasting models, seasonal pattern decomposition, markdown optimization, price elasticity modeling)
• Workflow Orchestration: N8N (purchasing recommendation generation, sell-through alert triggers, reorder workflows, overstock early warning notifications)
• Analytics & Dashboards: Power BI (sell-through tracking dashboards, location-level demand comparison, inventory aging reports, forecast-vs-actual variance analysis)
• Data Pipeline: Azure Data Factory (POS data extraction, cleaning, and consolidation)
• Database: Azure SQL (unified transaction history, forecast output storage)
• Integrations: 3 POS system data feeds, vendor order management systems