Almost every method in statistics assumes the rows could be shuffled without losing anything. Time series is the family where that's false. Yesterday's value informs today's; last December looks like this December. The order carries the information, and analyzing it as if the rows were independent throws that information away.
The payoff is forecasting — extending an observed pattern into the future, as the dashed continuation above suggests. But before you can forecast, you have to understand what the series is made of.
Note
Decomposition: The Parts of a Series
Every series is a sum of a few components. Separating them is the first step and often the most illuminating.
| Component | What it is |
|---|---|
| Trend | The long-run direction — up, down, or flat |
| Seasonality | A repeating cycle — weekly, monthly, yearly |
| Residual / noise | What's left after trend and season are removed |
A sales series that looks chaotic often resolves into a steady upward trend, a strong December spike, and a little noise. Trend analysis and seasonal decomposition make that structure explicit.
Stationarity: The Precondition
Many time series models require a stationary series — one whose statistical properties don't drift over time. A series with a trend or growing variance isn't stationary, and must usually be transformed (often by differencing) before modeling.
Watch out
The Forecasting Models
| Model | Best for |
|---|---|
| Exponential smoothing | Trend and seasonality, simple and robust |
| ARIMA | Autocorrelated series after differencing to stationarity |
| VAR / VECM | Several series that move together |
| GARCH family | Volatility that clusters — finance and risk |
ARIMA is the classic workhorse — it models a stationary series from its own past values and errors. Exponential smoothing is often just as good and simpler. VAR handles several interacting series at once; GARCH targets changing volatility rather than the level.
The One Rule: Never Shuffle Time
The most damaging mistake in time series is validating like ordinary data. Randomly splitting the rows lets the model peek at the future to predict the past — a leak that makes a useless model look excellent.
- 1Always train on the past and validate on the future, never a random split.
- 2Use a forward-chaining (rolling) split so every test window comes after its training window.
- 3Judge a forecast on data that genuinely came later, the way production will.
Tip
Time Series in the SKARI Statistical Lab
SKARI's Statistical Lab treats time series as its own family — the largest set of methods — with decomposition, stationarity checks, and forecasting built in.
- Trend analysis and seasonal decomposition
- Exponential smoothing and ARIMA forecasting
- VAR / VECM for multiple interacting series
- GARCH-family models for volatility
- Time-aware validation that never shuffles the order
Takeaway
Frequently Asked Questions
ARIMA or exponential smoothing?
Exponential smoothing is simpler and often competitive for trend-and-season data; ARIMA is more flexible for autocorrelated series. Try both and compare on a forward validation.
What is stationarity, quickly?
A series whose mean and variance don't drift over time. Trends and growing spread break it; differencing usually restores it.
Can I use a normal train/test split?
No — never shuffle time. Train on the past and test on the future, or the score is inflated by leakage.
Key Takeaways
Signal
Order
autocorrelation
Split into
Trend+Season
decompose
Needs
Stationary
difference first
Never
Shuffle
time moves forward
Time series analysis treats order as the signal, not a nuisance. Decompose the series, make it stationary, fit the model that matches its structure, and validate forward — and the forecast holds up when the real future arrives.
Takeaway
Predictive Modeling
Forecasting's supervised cousin
Cross-Validation
The time-aware split
Exploratory Data Analysis
Spot trend and season first