Rolling Windows in Polars — Finance Tutorial

A rolling window is a fixed-width slice that walks down the time series, recomputing a statistic at every step. Twenty-day moving average, twenty-day rolling volatility, Bollinger bands — every chart you have ever seen on a Bloomberg terminal is a rolling-window computation. In Polars, the entire family is one method on the expression — pl.col Close dot rolling_mean window_size twenty, paired with dot over Ticker for per-group windows. Source code: https://github.com/GoCelesteAI/polars-for-finance If returns are the language of finance from Episode 3, rolling windows are the grammar. This episode shows the substitution for every pandas groupby plus rolling pattern. Same fourteen-ticker dataset, three rolling expressions in one with_columns call, then arithmetic on the derived columns in a second with_columns call to build Bollinger bands. The null prefix on the first nineteen rows of each ticker proves the per-group window ran correctly. What You'll Build: - rolling.py — twenty-day simple moving average, fifty-day SMA, twenty-day rolling standard deviation, and Bollinger upper and lower bands, all per ticker, in two chained with_columns calls. - The rolling_mean idiom — pl.col Close dot rolling_mean window_size equals twenty dot over Ticker. Same shape as pct_change from Episode 3 — a windowed expression that aligns one-to-one with the input frame. - The two-stage with_columns pattern. The first stage builds rolling primitives from the raw Close. The second stage references the first stage's outputs to build derived bands. You cannot reference a newly created column within the same with_columns call. - The null-prefix sanity check. With window_size twenty and fourteen tickers, the rolling column has exactly two hundred and sixty six nulls — nineteen times fourteen. If your count is different, the dot over clause is missing and the window crossed ticker boundaries. - The full rolling family — rolling_mean, rolling_std, rolling_min, rolling_max, rolling_sum, rolling_median, rolling_quantile, rolling_skew — all take the same window_size argument and compose with dot over the same way. - The time-based variants — rolling_mean_by Date window_size five-d. Includes every row within the last five calendar days, regardless of how many trading bars that covers. Useful for data with holes. Timestamps: 0:00 - Intro — rolling windows, the grammar of finance 0:18 - Preview — three rolling stats plus Bollinger bands 0:54 - Open rolling.py in nvim 1:14 - Sort by Ticker and Date, then read the parquet 1:34 - rolling_mean and rolling_std with .over Ticker 2:00 - Second with_columns for Bollinger upper and lower bands 2:30 - Filter to AAPL and print the last eight rows 2:50 - Save and run 3:10 - Apple closing 273, SMA 20 at 276, bands 268 to 286 3:36 - End screen — recap and what's next Key Takeaways: 1. pl.col Close dot rolling_mean(20) dot over Ticker is the substitution for every pandas groupby plus rolling pattern. One line, one engine pass. The .over clause partitions by ticker and stitches the result back into the original row order. 2. Two with_columns calls. Stage one builds rolling primitives from the raw Close — sma twenty, sma fifty, std twenty. Stage two uses those primitives to build derived columns like Bollinger bands. You cannot reference a newly created column in the same with_columns call. 3. The null prefix is your proof. With window twenty and fourteen tickers, the rolling column has exactly two hundred sixty six nulls — nineteen times fourteen. Count them. If the count is wrong, dot over Ticker is missing and the rolling crossed ticker boundaries. 4. The full family covers every indicator you might want. rolling_mean, rolling_std, rolling_min, rolling_max, rolling_sum, rolling_median, rolling_quantile, rolling_skew. Same window_size argument, same compose-with-over pattern. 5. The rolling_underscore_by variants give you calendar-time windows instead of row-count windows. rolling_mean_by Date window_size five-d includes every row within the last five calendar days. Useful when your data has weekend or holiday holes. This channel is run by Claude AI. Tutorials AI-produced; reviewed and published by Codegiz. Source code at codegiz.com. #Polars #Python #Finance #DataAnalytics #RollingWindow #MovingAverage #BollingerBands #PythonForFinance #PolarsForFinance #TechnicalAnalysis --- Generated by Claude AI · part of the Polars for Finance series