📈 Lag Features, Rolling Windows & Cyclical Encoding

About this simulation

Tree-based and other non-sequential models (like XGBoost or random forests) have no built-in notion of "yesterday" or "last week" — every row is scored independently. To let them exploit time, engineers copy information across time into new columns: lag features (the value N steps ago), rolling statistics (mean/std over a trailing window), and cyclical encodings that map a periodic clock (hour, weekday, month) onto a circle via sine and cosine so "23:00" and "00:00" end up numerically close instead of far apart.

How it works

• The glowing sphere is the current row being scored, sliding along a synthetic seasonal-plus-noise signal.
• Orange vertical bars behind it mark lag features — copies of the value from 1…N steps earlier, wired into the current row as new input columns.
• The translucent box is the rolling window: the trailing N points averaged into a rolling mean, with an optional ±1σ band showing rolling volatility.
• The rotating wheel encodes the current time step as an angle around a circle, then reads off its sin (vertical) and cos (horizontal) projections — the two engineered columns that replace a single "raw hour" integer.

Controls

• Lag steps — how many lag(1)…lag(N) features are wired to the current row.
• Rolling window — trailing window size N for the rolling mean/σ band.
• Playback speed — how fast the current row advances through the series (0 = paused).
• Cyclical period — switches the wheel between hour-of-day, day-of-week and month-of-year encodings.
• Show rolling mean ± σ band — toggles the translucent volatility ribbon.
• ↺ New series — regenerates the synthetic signal with fresh noise and seasonality.

Did you know?

Building lag or rolling features on the full dataset before splitting into train/test — instead of computing them strictly from the past at each row — is one of the most common sources of data leakage in time-series modelling, producing validation scores that collapse in production.