Case Study Artifact Pipeline

Ad-hoc notebooks that load data, compute features, train models, and backtest in one file are impossible to debug, reproduce, or extend. This skill is about artifact boundaries and rerun rules, not about whether the research thesis is good.

The Problem

A quant writes a 500-line notebook that downloads data, engineers features, trains a model, and runs a backtest. It works once. Then: the data source changes, a feature is added, the model is retrained with different parameters, and the backtest uses stale predictions from the old model. Nobody can tell which outputs correspond to which inputs. The notebook becomes untouchable - too risky to change, too opaque to trust.

The Pattern

WRONG

# Monolithic notebook - everything in one file, no artifact boundaries
import polars as pl
from sklearn.linear_model import Ridge