Share of Voice forecasting system, Fortune 500 APAC (Artefact)

A Fortune 500 beauty company needed forecasts of Share of Voice, its share of media activity. The work covered 6 APAC markets, 5 product groups, and 4 divisions. It also covered Meta, Instagram, TikTok, and YouTube. Media leaders used monthly spreadsheets to find brands receiving too much or too little funding. However, these reports arrived 3 to 4 weeks late.

The data contained hundreds of short series with few observations. Each series covered a category, country, division, and platform. Campaign cycles created seasonal patterns, while competitor spending mixed with organic brand growth. Moreover, linear regression missed complex effects from platform changes and spending surges. A single forecast gave planners too little information about uncertainty. Therefore, they needed reliable ranges they could defend to finance leaders.

I replaced linear regression with XGBoost, which combines decision trees to capture complex patterns. Then I used SHAP to explain how each input affected the forecast. Client media teams checked these explanations against their knowledge. Moreover, BigQuery and dbt combined platform engagement data with SimilarWeb traffic and Traackr influencer signals.

I added split conformal prediction to XGBoost. This method uses held-out examples to set forecast ranges with coverage guarantees. Therefore, planners received intervals that reflected uncertainty alongside each forecast. The evaluation pipeline checked changes in these ranges and their observed coverage on rolling holdouts. It targeted 90% interval coverage alongside forecast accuracy.

I built the pipeline on Vertex AI Workbench. It automated data collection, versioned dbt features, tracked experiments, and supplied Streamlit dashboards. These dashboards turned model output into market-share views for commercial teams. Moreover, I was the office’s sole technical contributor. Therefore, I set standards for code review, repeatable results, and configuration.

I also built a proof of concept using retrieval-augmented generation, which supplies retrieved information to a language model. Business users could use it, while human reviewers checked factual accuracy and relevance. Moreover, LangSmith tracked prompt versions and responses.