Smooth a daily series without inventing observations or forecasts

Preserve the actual daily unit and grid

Choose a series that genuinely represents comparable daily observations, with timezone, period boundary and source meaning recorded. Three post rows may span two days or a month; averaging them is a three-observation summary, not automatically a three-day average.

Keep observed zero and missing distinct. If the source is cumulative, establish comparable interval changes first using the appropriate counter workflow. Do not smooth lifetime snapshots as if they were independent daily events or fill a source gap with an invented zero.

Choose the window and placement before reading the trend

A trailing three-day average at day t uses days t−2,t−1,t. A centered three-day average at t uses t−1,t,t+1. The centered value needs the later observation and cannot be known at t without an explicitly different forecasting assumption. NIST distinguishes these placements and the extra centering needed for even-sized windows.

Choose a window for an interpretable task, not whichever line looks smoothest after seeing results. A wider trailing window can hide a recent change. A seven-day average may summarize weekday variation, but is not automatically a fully seasonally adjusted series or proof that a recurring pattern is stable.

Read a complete constructed series and its lag

Constructed daily values are 10,20,30,40,50. The three-day trailing averages at days 3,4,5 are 20,30,40. The same three averages belong to days 2,3,4 when centered. Keep the first two trailing cells and the first/last centered cells unavailable under this declared full-window rule.

At day 5 the raw count is 50 but the trailing average is 40. That does not mean the original observation fell or that a method found ten missing events. Dallas Fed’s explanation emphasizes the trade-off between smoothing volatility and preserving the timeliness of a changing series.

Resolve three constructed smoothing decisions

S1 — Use 10,20,30,40,50 and the declared trailing window. Decision: publish the raw series with unavailable/unavailable/20/30/40, clearly labeled three-day trailing average. Outcome: a reproducible descriptive smoothing, not a forecast of day 6 or measured improvement in content.

S2 — Constructed missing day:10,20,unknown,40,50. Every three-day window shown intersects the unknown third day. Decision: leave those full-window averages unavailable rather than divide two observed days by three or silently skip the calendar gap. Outcome: no invented continuous daily series.

S3 — Constructed backtest mistake: a centered average assigned to day 2 includes day 3 and is presented as a prediction available on day 2. Decision: remove the real-time prediction claim or use an actually prior-only procedure. Outcome: correct historical placement without leaking future observations into an evaluation.

Keep overlap, revisions and competing explanations

Adjacent moving averages share observations. They are not a fresh independent sample at every plotted date, and a visually stable line does not justify treating each point as an independent trial. Keep the selected window, alignment and missing-value rule with any later analysis.

A revised source can change several smoothed points because it enters several windows. Preserve the source vintage and recompute transparently. Changes in publishing volume, collection or distribution can explain a pattern; smoothing cannot isolate a content effect or recover a full seasonal model from a few dates.

Save the grid and exact window recipe

The worksheet retains source/unit, dated raw grid, zeros/missing values, window size, alignment, complete-window rule, resulting series and permitted conclusion. The packet reproduces the small constructed arrays without a network or real analytics source.

Primary methodological pages were read on 1 November 2026. No real trend, forecast, audience or account result was measured. Use the separate prediction guide before evaluating any claim about future performance.