A feature baseline is a small, working prediction setup: choose an initial set of input features, pair it with a comparatively simple model, and measure the result before adding complexity. It gives the system a usable starting point and a reference for later changes. The aim is not to find the best possible feature set immediately; it is to learn whether additional information earns its implementation and maintenance cost. section 8.1 section 8.5
Build and extend the baseline in small steps:
- Start with a constant prediction or another simple working solution. Check whether the system produces predictions in the required format and establish an initial metric. section 8.1 section 8.2
- Select a basic feature set and train a simple model, such as linear regression. Add candidate features or transformations separately enough to see what each change contributes. section 8.6.1 section 8.5
- Compare metric gains with development effort, interpretability, and computation time. Stop adding complexity when the improvement no longer justifies its cost. section 8.5 section 8.7
Demand forecasting
For grocery demand forecasting, a first prediction can reuse sales from the same item and shop one week earlier to account for weekly seasonality. A linear-model baseline can then use sales lags and rolling summaries—such as sums, averages, medians, or quantiles over several windows. Candidate features also include price, revenue, customer counts, product category, and shop geography. Comparing these stages shows whether the extra inputs improve forecasts enough to warrant keeping them. section 8.6.1 section 8.5
Search relevance
A search engine can compare its existing keyword-based search with a simple ML baseline: turn queries and image metadata into TF-IDF features, then use a linear model to predict relevance. This tests an ML approach without first taking on image inputs or a more elaborate model. section 8.6.2
Keep the baseline during error analysis. Comparing a more complex model with simple predictions helps determine whether modeling and feature engineering are moving in the right direction, rather than merely adding complexity. section 9.4.1
