In signal processing and time-series analysis, a mathematical function applied to a segment of data to reduce artifacts at the segment edges before analysis, and in SQL and analytics, a calculation that operates over a sliding window of rows to compute running totals, rankings, and moving averages.
Also known as windowing function, sliding window function, SQL window function
Window function has two related but distinct meanings depending on context. In signal processing and audio analysis, a window function is a mathematical function applied to a segment of a signal before performing frequency analysis. When analyzing a finite-length segment of a continuous signal, sharp edges at the segment boundaries create artificial high-frequency artifacts (spectral leakage) in the frequency analysis. A window function tapers the signal smoothly to zero at both edges, reducing these artifacts and improving the accuracy of frequency estimates. Common window functions include Hann, Hamming, and Blackman windows.
In data analytics and SQL, a window function is a computation performed over a sliding ‘window’ of rows defined relative to each current row, without collapsing rows into groups the way GROUP BY does. SQL window functions can compute running totals (cumulative sum), moving averages, rankings within partitions, and row-to-row comparisons like period-over-period change. The OVER() clause defines the window—how rows are partitioned, ordered, and how far the window extends before and after each row. Window functions are fundamental to analytics SQL and are supported in all major databases and data warehouses.
In machine learning, a sliding window approach is used for time-series feature engineering: for each point in time, features are computed from the preceding N time steps (the window), allowing the model to incorporate recent context without encoding the full history. This is distinct from the signal processing usage but shares the core concept of analyzing a fixed-size segment of sequential data. The choice of window size is a hyperparameter that affects how much historical context the model can use.
Window functions in the SQL sense are among the most important analytical tools for agency campaign reporting. Moving averages smooth week-to-week performance fluctuations to reveal trends, period-over-period comparisons quantify growth rates, and cumulative spend calculations track budget pacing. Every major BI and analytics platform used in agency work—Looker, Tableau, BigQuery, Redshift—supports window functions, and AI-powered analytics tools that generate automated insights often rely on window function calculations as core building blocks.
Automated insight generation uses window functions at its core. Many AI-powered dashboard and reporting tools that agencies use for campaign analysis are built on SQL window function calculations: they compute 7-day moving averages, identify performance anomalies by comparing current values to rolling baselines, rank channels or creatives by cumulative contribution. Understanding what window functions are helps agencies evaluate whether an automated insight tool’s underlying calculations match their analytical intent and whether the window size parameters are appropriate for their reporting cadence.
Time-series AI models depend on window-based feature engineering. When AI tools predict future campaign performance from historical data, they typically use window-based feature engineering: for each prediction point, they extract features from the preceding days, weeks, or months of data. The choice of window size affects what patterns the model can learn. A model with a 7-day lookback window can learn day-of-week patterns but cannot capture monthly seasonality. Understanding this helps agencies specify appropriate window parameters when configuring or evaluating time-series prediction tools.
An agency’s data team is building automated weekly performance reports for a client with 12 DTC brands. The reports need to show each metric with a 4-week moving average alongside weekly actuals, and flag any week where performance drops more than 15% below the moving average. They implement this using SQL window functions in BigQuery: a 28-day rolling average using AVG() OVER(PARTITION BY brand ORDER BY week ROWS BETWEEN 27 PRECEDING AND CURRENT ROW), and a LAG() function to compute week-over-week change. An AI insight generation layer then uses these windowed metrics to identify which brands are underperforming relative to their own recent trend and draft narrative summaries. The window function layer makes the AI insights interpretable and auditable—the team can inspect exactly how each baseline and comparison was computed.
The workshop covers how AI tools actually work, how to evaluate them, and how to apply them to real agency workflows.