Case Study Development Workflow

A case study is not just a pipeline; it is a sequence of research decisions. Clean artifacts are necessary, but the real question is whether each stage earned the right to move to the next one.

The Problem

A researcher puts data loading, feature engineering, model training, and backtesting in a single notebook. It runs once and produces a Sharpe ratio. Later, they change the label horizon and must re-run everything. The model section silently reads stale features. The backtest uses predictions from a previous run. Nobody knows which config produced the final result. The pipeline is unreproducible and the results are untrustworthy.

The Pattern

WRONG

import lightgbm as lgb
import polars as pl

# Everything in one notebook - no artifact boundaries, no config
data = pl.read_parquet("etf_data.parquet")
features = data.select(["momentum_12m", "volatility_20d"])
labels = data["close"].pct_change(21).shift(-21)