Compute Daily Returns in Polars — Finance Tutorial

Daily returns are the language of finance. Every alpha model, every risk metric, every backtest opens with the same operation — take the close column, divide today by yesterday, subtract one. Trivial on a single ticker. The interesting question is what happens when fourteen tickers stack in the same frame. Source code: https://github.com/GoCelesteAI/polars-for-finance The answer in Polars is the .over clause — a window function that runs any expression per group while keeping the result aligned to the original row order. pl.col Close pct_change over Ticker computes six years of daily returns for fourteen tickers in one line, in parallel, with a query optimizer behind it. SQL window function semantics, Python syntax, Rust execution. This is the episode where the Polars expression API meets the bookkeeping work that powers every finance pipeline you will ever write. What You'll Build: - returns.py — sort the cached prices by Ticker and Date, then add daily_ret and log_ret columns in a single with_columns call. Fourteen-ticker frame, twenty eight thousand rows, two new columns computed in parallel. - The .over Ticker window pattern — append .over to any expression and Polars partitions, runs, and stitches the result back into the original frame. Same shape as SQL OVER PARTITION BY. - The log return variant — pl.col Close .log minus pl.col Close .log .shift one .over Ticker. The version you want for cumulative math because log returns add across periods. - The proof: count nulls in daily_ret. Fourteen nulls means the window ran per group. One null means you forgot the .over clause and the returns are silently garbage at every ticker boundary. - The .over composition pattern — rolling means, cumulative max for drawdown, ranks within ticker. Same shape, different inner expression. The window clause unlocks every windowed operation in the library. Timestamps: 0:00 - Intro — returns, the language of finance 0:18 - Preview — pct_change in one line for 14 tickers 1:00 - Open returns.py in nvim 1:18 - Sort by Ticker and Date 1:36 - daily_ret and log_ret with .over Ticker 2:02 - The null count proof 2:24 - Save and run 2:42 - Fourteen nulls — proof the window ran per group 3:10 - End screen — recap and what's next Key Takeaways: 1. .over Ticker is the window function clause. Append it to any expression and Polars runs the operation per group, keeping the output aligned to the original frame's row order. The semantics match SQL OVER PARTITION BY. This is the substitution that replaces every pandas groupby pct_change pattern in a finance pipeline, and it composes with filter, select, and with_columns the same way every other Polars expression does. 2. Sort first. The frame must be ordered by Ticker, Date for pct_change to mean anything. Polars does not auto-sort under .over — it trusts your input. The sort is a one-time twenty thousand row operation that takes milliseconds. Make it the first transformation after every read_parquet. 3. The null count is your proof the window ran correctly. Fourteen tickers means fourteen first-day nulls in the daily_ret column. If you see one null instead of fourteen, the .over clause is missing and you're computing returns across ticker boundaries — silently garbage, the same bug pandas has, neither library catches it for you. 4. Log returns are the variant for cumulative math. pl.col Close .log minus pl.col Close .log .shift one .over Ticker. Log returns sum across periods where simple returns compound. Build both at the same time — the marginal cost is zero and the right metric for each downstream calculation is always available. 5. .over composes with every window-shaped expression. Rolling mean over Ticker for moving averages. Cumulative max over Ticker for running peaks and drawdown. Rank over Ticker for within-ticker ordering. The library has fifteen plus window-capable methods; learn .over once and the rest is substitution. Episode 4 pairs this with group_by for the aggregation shape — same dataset, different output cardinality. This channel is run by Claude AI. Tutorials AI-produced; reviewed and published by Codegiz. Source code at codegiz.com. #Polars #Python #Finance #DataAnalytics #Returns #WindowFunctions #DataFrame #PythonForFinance #PolarsForFinance #PctChange --- Generated by Claude AI · part of the Polars for Finance series