Time Series Forecasting
summary
GumGum processes ~30 billion ad requests and 25+ TB of raw auction data per day and builds a real-time inventory forecasting pipeline that returns results under 30 seconds. The daily Spark pipeline samples data using an augmented MinHash approach to reduce 365 days of data to small Delta tables (~1.5 GB), caches them with Delta Lake and Delta caching, and runs forecasting in Spark with Auto ARIMA plus sinusoids, achieving consistently <5% MAPE for 30-day forecasts.
AI-written summary · about
chapters (9)
- 0:00Introduction and speakers
- 1:47About GumGum and use case
- 4:14Programmatic inventory forecast
- 7:26Forecasting architecture overview
- 9:31Data sampling approach
- 12:42Search and forecast application
- 16:18Delta Lake caching benefits
- 23:12Forecasting model and accuracy
- 25:21Q&A session
related talks
28mDemand Forecasting: Accuracy, Stability, Automation
29mProbabilistic demand forecasting at scale
40mMulti-Series Forecasting at Scale with StatsForecast
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
39mProductionising Data Science at Scale
18mUsing Transformer Models to Predict Demand at Scale in Production
17mOptimizing Data-Driven Decisions
56mSpark-driven Audience counting