Executive Summary:

Our client, a B2B software platform helping brands manage Amazon inventory, required higher forecast accuracy than Amazon's native Vendor Central could provide. InXiteOut developed a cross-channel machine learning solution that enriched historical sales data with web behavior, pricing, reviews, and promotional signals. By moving beyond isolated sales history, the custom model delivered a ~15% improvement in sell-through accuracy and produced decision-ready 26-week sell-in recommendations, enabling major vendors to optimize their inventory planning and purchase order execution.

Client Context

Our client, a B2B software platform that helps Amazon brands optimize their e-commerce performance, was consistently experiencing low accuracy with Amazon's native Vendor Central forecasts for inventory planning and sell-through predictions. Although the client actively curated a rich dataset of cross-channel metrics, including campaign responses, web behavior, online reviews, and pricing dynamics, this data remained siloed and was not being utilized to improve demand predictions.

They needed a machine learning solution to integrate these diverse signals with historical sales data, generating highly accurate sell-through forecasts and decision-ready sell-in recommendations for their vendors.

The InXiteOut Approach

We engineered an end-to-end forecasting pipeline that integrated siloed data streams into a unified predictive model. The solution was developed in three core phases:

Strategic Prioritization and Domain Mapping

We collaborated with the client to define baseline metrics and identify the specific parameters most likely to impact demand. From there, we established a comprehensive data enrichment strategy, mapping these attributes across the client’s existing, siloed data tables. To maximize business impact, we applied an ABC prioritization model to isolate high-value vendors and ASINs based on core revenue and volume KPIs.

Cross-Channel Data Enrichment

We constructed an enriched dataset to feed the machine learning pipeline by consolidating signals across five distinct categories:

• Web Behavior: Glance views, conversion rates, and lost buy box %.
• Online Reviews: Relative negative review metrics.
• Promotions: AMS spend, Prime Day indicators, and other special event flags.
• Pricing: Average buy box price, price discounts, and best deal indicators.
• Historical Sales: Past sell-through volumes and out-of-stock records.

Predictive Modeling and Sell-In Generation

We applied business-aware feature engineering to the enriched dataset to capture complex market dynamics like seasonality, promotional spikes, and price elasticity. A custom machine learning model was then trained and validated to forecast granular sell-through demand for each target vendor-ASIN combination.

Rather than stopping at raw demand predictions, the system combined these forecasts with current inventory levels and historical out-of-stock data. This integration allowed the tool to generate decision-ready sell-in recommendations over a 26-week lookahead horizon, equipping vendors with actionable insights for long-term supply chain planning and purchase order execution.

Technology Stack

• AWS ETL & Machine Learning Stack (Glue, S3, SageMaker, Redshift)

Benefits Delivered

The ML-driven forecasting solution was successfully integrated by the client, replacing Amazon’s native forecasts as the primary basis for inventory planning for their major vendors:

• ~15% Higher Forecast Accuracy: The custom sell-through predictions consistently outperformed Amazon Vendor Central baselines across all targeted vendor-ASIN combinations.
• High Vendor Adoption: The decision-ready 26-week sell-in recommendations were quickly adopted by the platform's top customers.
• Optimized Inventory Planning: Vendors achieved significantly better purchase order fulfillment outcomes by relying on enriched, cross-channel data rather than isolated sales history