Time Series Data Prep in Driverless AI
- Predictive AI
- H2O Driverless AI
This lesson walks through how to prepare time series data for H2O Driverless AI, covering the minimum two column structure of date and target, and how to handle multiple or grouped series such as sales by store and region. It covers best practices for enriching a dataset with external variables like holidays and pricing, sizing training data at least three times the test period, and matching test length to the forecast horizon. It also shows how to configure the time column and split a dataset by date directly in Driverless AI so training and test data stay in the correct order.
Key takeaways
- Why a time series dataset needs at minimum a date column and a target column to run in Driverless AI
- How grouped series, like sales by store or department, should be clustered before training a single experiment
- The training dataset should hold at least three times as much history as the test dataset and forecast horizon
- Why older historical data can hurt accuracy and should be trimmed to the relevant time window
- How to split training and test data by date using the split by time wizard in Driverless AI
Start owning your AI