Time-Series Forecasting
Predict the metric the business plans on
Forecasting predicts future values of a metric from its history, separating trend, seasonality, and noise. It turns calendar-aware data into a planning number that finance and ops can budget against.
What you will learn
- Decompose a series into trend, seasonality, and noise
- Build and evaluate a simple baseline and an additive forecast
- Validate forecasts out-of-sample and communicate uncertainty
What is this?
Time-series forecasting models a metric (revenue, DAU, bookings) as a function of its own past. The classic decomposition splits a series into trend (the long-run direction), seasonality (calendar patterns), and noise (unpredictable remainder). Forecasts extrapolate the pattern forward and are always paired with a band of uncertainty.
Why do you need it?
- Finance and operations budget against forecast numbers, not hopes.
- A naive baseline beats sophisticated models when the pattern is weak — so you need both.
- Understanding seasonality prevents misreading normal cycles as product changes.
Mental model
Estimating the next note in a song
Forecasting the next value of a series
You rely on the tune (trend), the rhythm (seasonality), and the fact that each note is close to the last one (autocorrelation).
How it works
- Resample the metric to a fixed frequency and visualize the last 1–2 years.
- Decompose into trend, seasonality (day-of-week, month-of-year), and residual noise.
- Fit a model — from a moving average baseline to ARIMA, Prophet, or lightGBM.
- Hold out the final period, compare forecast to actual, and report a confidence band.
Decomposing a series isolates the predictable pattern from the noise.
In production
- Finance forecasts monthly revenue with seasonality and a holiday uplift.
- Ops forecasts support volume to staff shifts ahead of cycle peaks.
- Growth detects product changes by comparing actuals to the forecast baseline.
Best practices
- Always beat a naive baseline (last-value or same-period-last-year) before trusting a model.
- Validate out-of-sample on the most recent, most business-relevant period.
- Report the forecast with its band — the business needs the range, not just the middle line.
Common mistakes
- Training on all history including a regime change, then predicting through the new regime.
- Stamping on a forecast because the actual was inside noise, not a real deviation.
- Ignoring calendar effects like holidays and weekdays in daily data.
Debugging tips
- If residuals show a repeating weekly spike, add the missing day-of-week feature.
- If the forecast follows a crash that already reversed, shorten the training window or add a change-point.
Interview checkpoint
Hands-on lab
- Decompose a daily revenue series into trend and day-of-week seasonality
- Build a seasonal-naive baseline and an additive model
- Evaluate both out-of-sample and report a confidence band
Forecast Revenue with Trend and Seasonality
45 min · advanced
Quick quiz
Time-series decomposition separates a series into
A forecast is only worth trusting when it